A SAR image denoising method based on space-frequency domain information aggregation

Abstract

The application provides a SAR image denoising method based on space-frequency domain information aggregation, uses a space-frequency domain information aggregation network architecture, the network architecture combines the features of the space domain and the frequency domain, extracts the space domain information and the frequency domain information of the image through a double-branch structure respectively, and fuses the two complementary features through a double-domain aggregation module, so that the speckle noise in the SAR image is effectively removed, and the image details are restored. Specifically, it includes a space domain information branch, a frequency domain information branch and a double-domain aggregation module. The space domain branch is used to capture local structure details, the frequency domain branch is used to extract global context information, and the double-domain aggregation module fuses the features from the space domain and the frequency domain to improve the speckle removal performance. The application is superior to the existing most advanced method under multiple noise levels and performs well in SAR image speckle removal. The application effectively combines local space features and global frequency features, thereby achieving better speckle removal effect.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and specifically to a SAR image denoising method based on spatial-frequency domain information aggregation. Background Technology

[0002] Synthetic Aperture Radar (SAR) is a microwave imaging radar technology that utilizes the relative motion between the antenna and the target, combined with advanced signal processing techniques, to synthesize a large aperture, thereby generating high-resolution radar images. Due to its all-weather, all-time operation capability, as well as its high spatial resolution and target recognition capabilities, SAR is widely used in military and civilian fields. However, SAR images are often affected by multiplicative speckle noise, which severely degrades image quality and negatively impacts subsequent processing tasks such as target detection, classification, and recognition. Figure 1 As shown. Therefore, the research on SAR image speckle removal algorithms has important theoretical and practical value.

[0003] Over the years, numerous image denoising algorithms have been developed, generally falling into two main categories: traditional denoising methods and deep learning-based methods. Traditional methods are further divided into three types: spatial domain filtering, transform domain filtering, and non-local mean (NLM) filtering. Methods such as Lee filtering, Kuan filtering, Frost filtering, and Gamma-MAP belong to spatial domain filtering. These methods reduce noise by utilizing the correlation of pixels in the spatial domain through sliding windows and statistical models. While these methods are intuitive, the choice of window size affects the denoising effect. Transform domain denoising methods first transform the image from the spatial domain to the transform domain through a specific transformation, then modify the coefficients in the transform domain, and finally reconstruct the denoised image through an inverse transform. Compared to spatial domain filtering, these methods perform better in preserving edge details, but the transformation process increases computational complexity. NLM filtering provides better denoising results by utilizing non-local similarities in the image. Significant advances in this field include the PPB filter and SAR-BM3D. These traditional methods achieve good denoising results in some scenarios, but their adaptability to complex scenarios is poor, and they often require manual parameter tuning to achieve optimal results.

[0004] In recent years, deep learning-based techniques have become the main tool for SAR image speckle removal. Chierchia et al. were the first to apply residual learning and batch normalization to SAR image denoising, using a network structure derived from Dn-CNN. This was a pioneering attempt, demonstrating the feasibility of convolutional neural networks in SAR image speckle removal. Zhang et al. proposed SAR-DRN, a SAR image speckle removal method that utilizes dilated residual networks (DRNs) to expand the receptive field, thereby improving denoising performance. Furthermore, some architectures have further improved denoising performance by introducing attention mechanisms to more efficiently capture image information. The aforementioned deep learning-based denoising methods have achieved good results in terms of visual quality and objective metrics. However, these methods mainly focus on the spatial domain, with limited exploration of potential solutions in the frequency domain.

[0005] Therefore, a method for re-examining SAR image speckle removal from the perspective of the frequency domain is needed. Summary of the Invention

[0006] This invention addresses the problem of poor speckle removal performance in SAR image denoising, particularly in high-noise or complex scenes. It provides a SAR image denoising method based on spatial-frequency domain information aggregation, employing a spatial-frequency domain information aggregation network architecture (SFSARNet). This architecture combines features from both the spatial and frequency domains. A dual-branch structure extracts spatial and frequency domain information from the image separately, and a dual-domain aggregation module fuses these complementary features, effectively removing speckle noise and restoring image details in SAR images. Specifically, it includes a spatial domain information branch, a frequency domain information branch, and a dual-domain aggregation module. The spatial domain branch captures local structural details, the frequency domain branch extracts global contextual information, and the dual-domain aggregation module fuses features from both the spatial and frequency domains to improve speckle removal performance. Experimental results show that this invention outperforms state-of-the-art methods at multiple noise levels, demonstrating excellent performance in SAR image speckle removal. This invention uses Fast Fourier Transform (FFT) to convert the spatial domain image into an amplitude-phase spectrum in the frequency domain, effectively combining local spatial features with global frequency features to achieve better speckle removal, especially in detail preservation and noise suppression.

[0007] This invention provides a SAR image denoising method based on spatial-frequency domain information aggregation, comprising the following steps:

[0008] S1. Input the SAR image to be denoised into the spatial domain information branch of the spatial-frequency domain information aggregation network architecture. The spatial domain information branch extracts features to obtain spatial features F through the U-Net network and the supervised attention module. spa Then, the spatial domain restored image I is obtained by decoding and recovery. s ;

[0009] S2. Input the SAR image to be denoised into the frequency domain information branch of the space-frequency domain information aggregation network architecture, obtain the frequency domain representation of the SAR image through Fourier transform, and convert the real and imaginary parts of the frequency domain representation of the SAR image into the amplitude of the SAR image to be denoised. and phase

[0010] amplitude Perform a logarithmic transformation to obtain the amplitude of the logarithmically transformed noise image;

[0011] The amplitude and phase of the noisy image after logarithmic transformation The first and second sub-branches of the frequency domain information branch are input respectively. The first sub-branch extracts the amplitude feature F from the logarithmically transformed noisy image through the U-Net network and the supervised attention module. amp Then, the recovered amplitude spectrum is obtained by decoding and restoring and performing an exponential transformation. The second sub-branch extracts phase using a U-Net network and a supervised attention module. The characteristics of the phase feature F are obtained. pha After decoding and recovery, the recovered phase spectrum is obtained.

[0012] Restore amplitude spectrum and recovering the phase spectrum The real and imaginary parts of the recovered Fourier domain image are converted respectively, and then converted back to the spatial domain by inverse Fourier transform to obtain the frequency domain branched recovered image I. f ;

[0013] S3, the dual-domain aggregation module of the space-frequency domain information aggregation network architecture will integrate the amplitude feature F amp and phase characteristics F pha Convert the parts to real and imaginary parts respectively, and then convert them back to the spatial domain using inverse Fourier transform to obtain the global representation feature F. fre ;

[0014] The dual-domain aggregation module uses an attention mechanism to process spatial features F spa and global representation features F fre The differences are then subjected to global average pooling to obtain attention weights, which are then combined with the spatial features F. spa After element-wise multiplication, it is combined with the spatial feature F. spa The weighted summation yields the enhanced local feature representation F' spa ;

[0015] The enhanced local feature representation F' spa With global representation features F fre The image I is recovered by splicing convolutional reconstruction. a ;

[0016] S4, the dual-domain aggregation module restores the spatial domain image I s Frequency domain branched image reconstruction I f Image recovery module and aggregation module a The summation yields a clear image I C ;

[0017] S5. Based on the loss function Training space-frequency domain information aggregation network architecture, loss function Including spatial domain loss and frequency domain loss After achieving the goal, a trained spatial-frequency domain information aggregation network architecture is obtained. The trained spatial-frequency domain information aggregation network architecture is then used for SAR image denoising, thus completing a SAR image denoising method based on spatial-frequency domain information aggregation.

[0018] The SAR image denoising method based on spatial-frequency domain information aggregation described in this invention, as a preferred embodiment, includes a U-Net network comprising residual blocks and dense connections in the spatial domain information branch. A supervised attention module is connected to the end of the U-Net network and uses the input information to adjust the feature map to obtain spatial features F. spa .

[0019] The SAR image denoising method based on spatial-frequency domain information aggregation described in this invention, as a preferred embodiment, in step S2, and They are respectively The real and imaginary parts, SAR image frequency domain representation, where x is the input image;

[0020]

[0021] in, For amplitude spectrum, Phase spectrum;

[0022]

[0023] in, and These represent the real and imaginary parts of the recovered Fourier domain image, respectively.

[0024] The SAR image denoising method based on spatial-frequency domain information aggregation described in this invention, as a preferred embodiment, in step S2, involves adjusting the amplitude... The logarithmic transformation converts multiplicative noise into additive noise; after the logarithmic transformation, the SAR speckle noise distribution becomes Gaussian.

[0025] In the SAR image denoising method based on spatial-frequency domain information aggregation described in this invention, as a preferred embodiment, in step S3, an enhanced local feature representation F' is obtained using a 1×1 convolution kernel. spa The enhanced local feature representation F' is obtained by using a convolutional branch with a 3×3 kernel. spa With global representation features F fre splicing.

[0026] The SAR image denoising method based on spatial-frequency domain information aggregation described in this invention, as a preferred embodiment, includes step S3,

[0027]

[0028] Where GAP represents global average pooling, ⊙ represents element-wise multiplication, and Conv... 1×1 This is a 1×1 convolution operation, where α is a learning scalar initialized to 1.

[0029] The SAR image denoising method based on spatial-frequency domain information aggregation described in this invention, as a preferred embodiment, is I C =I s +I f +I a .

[0030] The SAR image denoising method based on spatial-frequency domain information aggregation described in this invention, as a preferred embodiment, includes step S4, λ is the weighting factor.

[0031] The SAR image denoising method based on spatial-frequency domain information aggregation described in this invention, as a preferred embodiment, I G For I C The corresponding real image;

[0032] in, and phase components They are respectively using Fast Fourier Transform to convert I G Amplitude after conversion to Fourier space And phase; and These are the recovered amplitude spectrum and the recovered phase spectrum obtained from the frequency domain information branch, respectively.

[0033] The SAR image denoising method based on spatial-frequency domain information aggregation described in this invention, as a preferred embodiment, in step S4, the training dataset of the spatial-frequency domain information aggregation network architecture includes synthetic SAR images and images with speckle noise. During training, speckle noise in SAR images is simulated by adding multiplicative noise with a Gamma distribution.

[0034] The performance of the spatial-frequency domain information aggregation network architecture was optimized by using the AdamW optimizer and adjusting the learning rate, batch size, and number of training iterations.

[0035] The training parameters included: batch size of 32, initial learning rate of 0.0005, and training for 300 epochs.

[0036] The present invention has the following advantages:

[0037] (1) This invention fully combines spatial domain and frequency domain information to make up for the shortcomings of traditional deep learning denoising networks that mainly rely on the single domain method of spatial domain for feature extraction. By deeply fusing spatial domain and frequency domain features, it not only uses the spatial domain branch to obtain fine-grained local features, but also combines the global information of the frequency domain branch, so that noise removal and detail restoration can be optimized in a coordinated manner.

[0038] (2) This invention employs an attention mechanism for adaptive feature fusion, dynamically adjusting the enhancement method of spatial features rather than simply concatenating or weighted summing. Specifically, the attention weights are adaptively calculated based on the relationship between spatial domain features and global frequency domain features, enabling local feature enhancement to be guided by global information, thus avoiding the problem of traditional CNNs losing overall structural information due to local dependencies during denoising.

[0039] (3) Unlike previous methods that only use the frequency domain as auxiliary information, this invention converts the noise reduction in the frequency domain back to the spatial domain and further supplements the image details through enhanced local features to form variational domain information. This ensures that the frequency domain noise reduction does not cause the loss of structural information. This cross-domain information interaction improves the robustness of the model and enables it to better adapt to SAR images with different noise levels.

[0040] (4) In traditional frequency domain denoising methods, Fourier transform often leads to loss of high-frequency information. However, this invention uses multi-scale feature fusion, which combines 1×1 convolution with 3×3 convolution branches to further restore local details in the spatial domain and can be deeply fused with frequency domain information, thereby preserving texture information while denoising and improving the final visual quality.

[0041] (5) This invention is the first to combine spatial domain and frequency domain information for SAR image despeccing, effectively utilizing the local and global features of the image. The network architecture processes spatial domain and frequency domain information separately through a dual-branch structure, and fuses the features of both through a dual-domain aggregation module, thereby significantly improving the despeccing effect. Experiments show that this invention surpasses the existing state-of-the-art methods under multiple noise conditions, and performs well in both despeccing effect and detail preservation. This method has good transferability and can be applied to other existing denoising networks. Attached Figure Description

[0042] Figure 1 This is a schematic diagram illustrating the impact of speckle noise on SAR images and their amplitude and phase spectra.

[0043] Figure 2 This is a flowchart of a SAR image denoising method based on spatial-frequency domain information aggregation;

[0044] Figure 3 This is a diagram of the spatial-frequency domain information aggregation architecture for a SAR image denoising method based on spatial-frequency domain information aggregation.

[0045] Figure 4a SAR1 is an example of a SAR image denoising method based on spatial-frequency domain information aggregation;

[0046] Figure 4b This is a SAR2 image from an embodiment of a SAR image denoising method based on spatial-frequency domain information aggregation;

[0047] Figure 4c This is a SAR3 image from an embodiment of a SAR image denoising method based on spatial-frequency domain information aggregation;

[0048] Figure 5 This is a comparison of the denoising results of an embodiment of a SAR image denoising method based on spatial-frequency domain information aggregation and a comparison network on SAR1 images.

[0049] Figure 6 This is a comparison of the denoising results of an embodiment of a SAR image denoising method based on spatial-frequency domain information aggregation and a comparison network on SAR2 images.

[0050] Figure 7 This is a comparison of the denoising results of an embodiment of a SAR image denoising method based on spatial-frequency domain information aggregation and a comparative network on SAR3 images. Detailed Implementation

[0051] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0052] Example 1

[0053] like Figure 2 , 3 As shown, a SAR image denoising method based on spatial-frequency domain information aggregation includes the following steps:

[0054] S1. Design the spatial domain information branch and recover the details of the spatial domain image through the U-Net architecture;

[0055] This branch uses the U-Net architecture to recover noisy images in the spatial domain. N To further enhance feature representation, the U-Net architecture employs residual blocks and dense connections. Furthermore, a supervised attention module (SAM) is introduced at the end of each branch to adjust the feature map using input information and generate the final spatial feature F. spa Ultimately, the spatial domain recovery result I can be obtained from the spatial domain branch. s Next, F spa and I s The intermediate result will be input into the two-domain aggregation module.

[0056] This embodiment recovers spatial domain details of an image using the U-Net architecture. The U-Net architecture, consisting of an encoder and a decoder, effectively captures local image features and preserves contextual information through skip connections. By using residual blocks and a supervised attention module (SAM), this embodiment further enhances the network's ability to express image features and recover more image details.

[0057] S2. Design a frequency domain information branch and recover the amplitude and phase information of the frequency domain image through Fourier transform;

[0058] Speckle noise in SAR images is often described as multiplicative noise. Let x be an image with speckle noise, X be a noise-free image, and N be the speckle noise components, then it is defined as:

[0059] x = N·X

[0060] It is usually assumed that N follows a Gamma distribution with mean 1 and variance 1 / L, and its probability density function is defined as:

[0061]

[0062] Where Γ(·) represents the Gamma function and L represents the number of looks.

[0063] The amplitude of a noisy image varies significantly with different noise levels, while the phase remains relatively stable because it encodes the image's structural information. Due to the non-additive nature of multiplicative noise, and its primary impact on amplitude rather than phase, direct denoising in the spatial domain may be insufficient to effectively suppress noise and could potentially damage the image's structural information. To address this issue, this invention first performs a Fourier transform on the image, decomposing it into an amplitude spectrum and a phase spectrum. Since the amplitude spectrum is significantly affected by multiplicative noise, this invention performs a logarithmic transform on it, converting the original multiplicative noise into additive noise. This makes subsequent denoising methods (such as filtering or neural networks) easier to model and process. This transformation reduces the non-uniformity caused by multiplicative noise, making the statistical distribution of the noise closer to a normal distribution. Since SAR speckle noise typically follows a Gamma distribution, after the logarithmic transform, it mathematically approximates a Gaussian distribution, making it easier for denoising methods (such as deep learning or adaptive filtering) to model the noise distribution, improving the robustness of the denoising network, while preserving crucial structural information. Based on this, the amplitude and phase information are recovered in the frequency domain using a frequency domain branch, and a clear image is reconstructed after applying an inverse transform.

[0064] In this embodiment, the frequency domain information branch extracts the amplitude and phase spectra of the image through Fourier transform. Then, the frequency domain information is recovered using the U-Net architecture.

[0065] By performing a Fourier transform on the noisy image, the frequency domain representation of the image is obtained as follows:

[0066]

[0067] Where x represents the input image, Indicates Fourier transform, and They represent The real and imaginary parts. These components are then converted into amplitude and phase spectra, as follows:

[0068]

[0069] in, Indicates amplitude spectrum, This represents the phase spectrum. The amplitude of the noisy image is obtained using this formula. and phase right Perform a logarithmic transformation to change the phase and logarithmic transformation The data is fed into two sub-branches for recovery to obtain phase and amplitude characteristics, i.e., F. pha and F ampThese two sub-branches, like the spatial domain branch, both employ the U-Net architecture and the SAM module. Similarly, according to the formula, F... pha Calculate the recovered phase By F amp Obtain the magnitude of the recovery and perform an exponential transformation to obtain... By substituting the recovered amplitude and phase into the following formula, the real and imaginary parts of the recovered Fourier domain image can be calculated:

[0070]

[0071] Finally, the recovered frequency domain is transformed back into the spatial domain through inverse Fourier transform to obtain a clear image I. f .

[0072] The recovered amplitude and phase are calculated using the following formula:

[0073]

[0074] here, Indicates the extent of recovery. This indicates the recovered phase.

[0075] S3. Design a dual-domain aggregation module to fuse features from the spatial domain and the frequency domain to generate a speckle-free image;

[0076] Convolution operations are highly effective at capturing local features, but can have limitations when processing global information. According to the spectral convolution theorem in Fourier theory, frequency domain information can be processed to capture a global representation. In short, the spatial domain branch captures the local representation of an image, while the frequency domain branch captures the global representation. The representations generated by these two branches are complementary, and effectively aggregating them helps to recover image details.

[0077] To effectively fuse spatial and frequency domain features, this invention employs an attention mechanism to adaptively enhance important spatial representations while considering global frequency information. Attention mechanisms are widely used in deep learning; they dynamically assign weights to different parts of the input, enabling the model to focus on the most relevant features while suppressing unimportant information. In this framework, the attention mechanism is used to modulate spatial features based on the difference between the spatial and frequency representations, ensuring that local feature enhancement is guided by the global context. The detailed steps of the dual-domain aggregation are as follows: First, the amplitude feature F... amp and phase characteristics F pha The data is converted to real and imaginary parts, and then transformed back to the spatial domain using the inverse Fourier transform to obtain the global representation feature F. fre Then, for F spa Further enhancements are performed to obtain a better local representation. This involves enhancing the local feature F'.spa It is expressed as a weighted sum of the original features and the features of interest, as shown in the following formula:

[0078] F' spa =Atten⊙Conv 1×1 (F spa )+αF spa

[0079] Atten = Softmax[GAP(F spa -F fre )]

[0080] Where GAP represents global average pooling, ⊙ represents element-wise multiplication, and Conv... 1×1 This represents a 1×1 convolution operation, where α is a learning scalar initialized to 1. Attention weights are obtained by performing global average pooling (GAP) on the difference between spatial domain features and global frequency domain features, followed by a softmax function. The result calculated through this attention mechanism is compared with the F-value obtained through the 1×1 convolution operation. spa Performing element-wise multiplication, the result is compared with the original feature F. spa Weighted summation yields the enhanced local feature representation F' spa Then, the present invention enhances the representation of local features and global features F. fre The data is then stitched together, and a 3×3 convolutional branch is used to further predict the recovered result I. a .

[0081] Finally, clear image I C Spatial domain branch recovery result I s Frequency domain branch recovery result I f Recovery results of aggregation module I a The sum of.

[0082] I C =I s +I f +I a

[0083] In this embodiment, the dual-domain aggregation module is responsible for weighting and fusing the features from the spatial and frequency domains. First, the recovered frequency domain image (amplitude and phase) is converted back to the spatial domain using inverse Fourier transform:

[0084]

[0085] Generate the final despiked image.

[0086] S4. Design a loss function and train the network to obtain the speckle-removed SAR image.

[0087] In the training process of this embodiment, the difference between the denoised image and the real image is calculated using L1 loss:

[0088]

[0089] Frequency domain loss: First, use Fast Fourier Transform to convert I... G Transform to Fourier space and calculate its magnitude. and phase components Next, calculate and L1 loss between, and and The L1 loss between them. Here, and It is obtained from the frequency domain branch. The total frequency domain loss is given by the following formula:

[0090]

[0091] The total loss function consists of a weighted sum of spatial domain loss and frequency domain loss, and the optimization objective of this embodiment is to minimize it:

[0092]

[0093] Where λ is the weighting factor, which is usually set to 0.1.

[0094] Optimization methods and parameter settings during training

[0095] During training, the AdamW optimizer is used to optimize the network. Specifically, the training hyperparameters are set as follows in this embodiment:

[0096] Batch size: 32

[0097] Initial learning rate: 0.0005

[0098] Training rounds: 300

[0099] The learning rate half-decays every 20 epochs.

[0100] During training, a multiplicative noise model with a Gamma distribution is used to simulate speckle noise in SAR images, and noise is added to sharp images to generate noisy image pairs for training. The network performs self-supervised learning using these noisy image pairs, without requiring clean images as targets.

[0101] After acquiring SAR images and converting them from the spatial domain to the frequency domain using Fourier transform, frequency domain features are recovered using the results of the Fourier transform.

[0102] In this step, the SAR image is converted from the spatial domain to the frequency domain. The amplitude and phase spectra are obtained through Fourier transform, and their amplitude and phase information is recovered using the U-Net architecture. The recovered frequency domain amplitude and phase are then converted back to the spatial domain through inverse Fourier transform, resulting in the final speckle-free image.

[0103] When designing the dual-domain aggregation module, the recovered frequency domain information is transformed back to the spatial domain using inverse Fourier transform. This spatial domain information is then weighted and fused with the recovered image features. The weighted fusion process employs soft max pooling and 1×1 convolution operations to weight and fuse the spatial and frequency domain features, thereby further enhancing the image restoration effect. The final speckle-free image is obtained by weighted synthesis of the restored spatial and frequency domain results.

[0104] Optionally, the loss function of the design includes:

[0105] Spatial domain loss: The difference between the denoised image and the real image is calculated using L1 loss;

[0106] Frequency domain loss: Optimizes denoising performance by calculating the difference between the recovered amplitude and phase and the amplitude and phase of the real image;

[0107] Total loss function: The spatial domain loss and frequency domain loss are weighted and summed to obtain the final loss function, which is used to train the network.

[0108] This embodiment conducted experiments on real SAR images. For comparison, several common denoising methods were selected, with Lee and SAR-BM3D as traditional methods, and SIFSDNet, MRDDANet, and SAR-CAM as deep learning-based methods.

[0109] In real SAR data experiments, due to the difficulty in obtaining noise-free reference images, this embodiment uses several common no-reference evaluation metrics to measure the performance of the method, including Equivalent Number of Looks (ENL), Image Mean (MoI), Mean Ratio (MoR), and Edge Preservation Based on Mean Ratio (EPD-ROA). EPD-ROA is evaluated in both the horizontal (HD) and vertical (VD) directions. A higher ENL value indicates better denoising performance, while the closer the results of the other metrics are to 1, the better the denoising effect. ENL, MoI, and MoR are mainly used to evaluate denoising performance in homogeneous regions, while EPD-ROA is mainly used to evaluate denoising performance in non-homogeneous regions (such as textures and edges).

[0110] like Figures 4a-4cAs shown, to verify the denoising effect of SFSARNet on real SAR images, this embodiment randomly selected three SAR images from different scenes. These three images are named SAR1, SAR2, and SAR3, respectively. Figures 4a-4c In the diagram, the red box represents the uniform region used to calculate ENL, MoI, and MoR, while the yellow box represents the edge feature region used to calculate EPD-ROA.

[0111] This embodiment uses different methods for testing and displays the denoised image results. Figure 5-7 middle.

[0112] from Figure 5 and Figure 6 It can be seen that Lee filtering has poor denoising effect. SAR-BM3D and MRDDANet have certain denoising capabilities, but texture details become blurred in magnified areas. SIFSDNet, SAR-CAM, and the method proposed in this invention perform better in edge texture preservation, achieving a balance between denoising and image detail preservation.

[0113] Figure 7 The visual results of SAR3 image denoising using the network of this invention compared to other contrasting networks are shown. From Figure 7 It can be seen that SAR-BM3D has strong denoising capabilities, but the denoised image is too smooth; Lee, SIFSDNet, and MRDDANet have poor denoising effects, and the denoised image still has a lot of speckle noise; SAR-CAM has some denoising capabilities, but some noise still remains in the magnified area. In comparison, this embodiment has the best visual effect after denoising.

[0114] To provide a more objective comparison, Table 1 presents the quantitative evaluation results of ENL, MoI, MoR, and EPD-ROA, with the best results indicated in bold and the second-best results indicated by underline. Figure 4a Table 1 shows the regions selected for different evaluations, where red rectangles represent uniform regions used to calculate ENL, MoI, and MoR, and yellow rectangles represent regions containing rich edge features used to calculate EPD-ROA. As can be seen from Table 1, this embodiment achieves the best performance in MoI, MoR, and EPD-ROA metrics, while also reaching the first or second level in ENL. This indicates that the present invention not only has strong denoising capabilities but also excellent detail recovery capabilities.

[0115] Furthermore, in terms of computational efficiency, this invention is slightly inferior to SAR-CAM, mainly because the introduction of Fourier transform and inverse Fourier transform in the network increases computation time. However, this invention has stronger denoising capabilities, achieving a better balance between denoising effect and detail preservation.

[0116] Table 1. Quantitative evaluation results of the real dataset

[0117]

[0118]

[0119] Compared with existing technologies, this invention combines features from the spatial and frequency domains to simultaneously optimize image denoising from both global and local perspectives. By designing a specialized loss function, it can effectively preserve the texture details and edge information of the image while denoising. This method can be transferred to other existing SAR despeculiarization networks and can be effectively applied to different types of SAR image denoising tasks.

[0120] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A SAR image denoising method based on spatial-frequency domain information aggregation, characterized in that: Includes the following steps: S1. The SAR image to be denoised is input into the spatial domain information branch of the spatial-frequency domain information aggregation network architecture. The spatial domain information branch extracts features to obtain spatial features F through a U-Net network and a supervised attention module. spa Then, the spatial domain restored image I is obtained by decoding and recovery. s ; S2. Input the SAR image to be denoised into the frequency domain information branch of the space-frequency domain information aggregation network architecture, obtain the frequency domain representation of the SAR image through Fourier transform, and convert the real and imaginary parts of the frequency domain representation of the SAR image into the amplitude of the SAR image to be denoised. and phase The magnitude Perform a logarithmic transformation to obtain the amplitude of the logarithmically transformed noise image; The amplitude and phase of the logarithmically transformed noise image The first and second sub-branches of the frequency domain information branch are input respectively. The first sub-branch extracts the amplitude feature F from the logarithmically transformed noisy image through a U-Net network and a supervised attention module. amp Then, the recovered amplitude spectrum is obtained by decoding and restoring and performing an exponential transformation. The second sub-branch extracts the phase through a U-Net network and a supervised attention module. The characteristics of the phase feature F are obtained. pha After decoding and recovery, the recovered phase spectrum is obtained. The recovered amplitude spectrum and the recovered phase spectrum The real and imaginary parts of the recovered Fourier domain image are converted respectively, and then converted back to the spatial domain by inverse Fourier transform to obtain the frequency domain branched recovered image I. f ; S3, the dual-domain aggregation module of the space-frequency domain information aggregation network architecture will aggregate the amplitude feature F amp and the phase feature F pha Convert the parts to real and imaginary parts respectively, and then convert them back to the spatial domain using inverse Fourier transform to obtain the global representation feature F. fre ; The dual-domain aggregation module uses an attention mechanism to process the spatial feature F. spa and the global representation feature F fre The differences are then subjected to global average pooling to obtain attention weights, which are then combined with the spatial feature F. spa After element-wise multiplication, it is combined with the spatial feature F. spa The weighted summation yields the enhanced local feature representation F' spa ; The enhanced local features are represented as F' spa With the global representation feature F fre The image I is recovered by splicing convolutional reconstruction. a ; S4. The dual-domain aggregation module restores the spatial domain image I. s The frequency domain branched image I f The aggregation module recovers image I a The summation yields a clear image I C ; S5. Based on the loss function The loss function is used to train the spatial-frequency domain information aggregation network architecture. Including spatial domain loss and frequency domain loss After achieving the goal, a trained spatial-frequency domain information aggregation network architecture is obtained. The trained spatial-frequency domain information aggregation network architecture is then used for SAR image denoising, thus completing a SAR image denoising method based on spatial-frequency domain information aggregation.

2. The SAR image denoising method based on spatial-frequency domain information aggregation according to claim 1, characterized in that: In the spatial domain information branch, the U-Net network includes residual blocks and dense connections. A supervised attention module is connected to the end of the U-Net network and uses the input information to adjust the feature map to obtain the spatial feature F. spa .

3. The SAR image denoising method based on spatial-frequency domain information aggregation according to claim 1, characterized in that: In step S2, and They are respectively The real and imaginary parts, Image frequency domain representation, where x is the input image; in, For amplitude spectrum, Phase spectrum; in, and These represent the real and imaginary parts of the recovered Fourier domain image, respectively.

4. The SAR image denoising method based on spatial-frequency domain information aggregation according to claim 1, characterized in that: In step S2, by the amplitude Performing a logarithm converts multiplicative noise into additive noise; After logarithmic transformation, the SAR speckle noise distribution becomes Gaussian.

5. The SAR image denoising method based on spatial-frequency domain information aggregation according to claim 1, characterized in that: In step S3, the enhanced local feature representation F' is obtained by using a convolution with a 1×1 kernel. spa The enhanced local feature representation F' is generated using a convolutional branch with a 3×3 kernel. spa With the global representation feature F fre splicing.

6. The SAR image denoising method based on spatial-frequency domain information aggregation according to claim 5, characterized in that: In step S3, Where GAP represents global average pooling, ⊙ represents element-wise multiplication, and Conv... 1×1 This is a 1×1 convolution operation, where α is a learning scalar initialized to 1.

7. The SAR image denoising method based on spatial-frequency domain information aggregation according to claim 1, characterized in that: I C =I s +I f +I a 。 8. The SAR image denoising method based on spatial-frequency domain information aggregation according to claim 1, characterized in that: In step S4, λ is the weighting factor.

9. A SAR image denoising method based on spatial-frequency domain information aggregation according to claim 8, characterized in that: I G For I C The corresponding real image; in, and phase components They are respectively using Fast Fourier Transform to convert I G Amplitude after conversion to Fourier space And phase; and These are the recovered amplitude spectrum and the recovered phase spectrum obtained from the frequency domain information branch, respectively.

10. A SAR image denoising method based on spatial-frequency domain information aggregation according to claim 1, characterized in that: In step S4, the training dataset of the space-frequency domain information aggregation network architecture includes synthetic SAR images and images with speckle noise. During training, multiplicative noise with a Gamma distribution is added to simulate speckle noise in the SAR images. The performance of the spatial-frequency domain information aggregation network architecture was optimized by using the AdamW optimizer and adjusting the learning rate, batch size, and number of training iterations. The training parameters included: batch size of 32, initial learning rate of 0.0005, and training for 300 epochs.

Images