SAR image denoising network method and system based on adaptive multi-scale feature fusion AMFFD-Net

Abstract

The present invention provides a SAR image denoising network method and system of adaptive multi-scale feature fusion AMFFD-Net, the method includes: data set preparation; data preprocessing; constructing an adaptive multi-scale feature fusion model AMFFD-Net, processing to obtain and output denoised SAR images; the adaptive multi-scale feature fusion model AMFFD-Net includes: a noise estimation module, an asymmetric convolution module, an adaptive feature selection module, a ReLU activation module, a multi-scale feature fusion module, an SE attention module, a cascade block and a deep residual learning module. The present invention solves the technical problems of insufficient detail and texture retention, blurred edges and structures, poor high-frequency information processing, limited adaptability and weak generalization ability.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision recognition and synthetic aperture radar (SAR) automatic target recognition, and in particular to a SAR image denoising network method and system of adaptive multi-scale feature fusion AMFFD-Net. Background Art

[0002] Synthetic aperture radar (SAR) is a high-resolution radar system widely used in fields such as earth observation, military reconnaissance and environmental monitoring. Compared with traditional optical imaging, SAR systems can acquire surface images under all-weather and all-day conditions, and have the advantage of not being affected by light and weather. However, speckle noise is inevitably introduced in the process of acquiring SAR images, which is caused by the superposition of the Doppler effect caused by the coherence of radar waves. Speckle noise can significantly reduce the image quality and cause difficulties in subsequent image processing and analysis.

[0003] To overcome this challenge, researchers have developed a variety of denoising techniques. Early methods included linear filters, such as average filters and Gaussian filters, which are simple and fast, but blur the image and lose details. Subsequently, nonlinear filters such as Lee filters, Frost filters, and Kuan filters have been widely used. They improve the denoising effect by considering the statistical characteristics of local areas. For example, the existing invention patent application document "SAR image edge detection method and device combined with wide line extraction" with publication number CN110796674A, the existing method includes: obtaining edge intensity information of the SAR image in all directions, constructing a preliminary edge intensity map, convolving the preliminary edge intensity map with a preset Gaussian kernel function, obtaining a Hessian matrix corresponding to the preliminary edge intensity map, obtaining an enhanced edge intensity map of the SAR image according to the eigenvalues ​​and eigenvectors of the Hessian matrix, filtering the enhanced edge intensity map with a preset isotropic nonlinear filter, obtaining wide lines in the enhanced edge intensity map, extracting the center line of the wide lines, and obtaining the SAR image edge. And the existing invention patent application document "A multi-angle SAR image filtering method based on super image" with publication number CN112927155A, the existing method includes: step 1), pre-processing the multi-angle SAR image and selecting a reference image; step 2), resampling other images based on the reference image; step 3), putting the reference image and the resampled image together for mean processing to obtain a super image, and performing enhanced Lee filtering on it; step 4), performing ratio processing on N-scene multi-angle SAR images and the filtered super image, and performing enhanced Lee filtering on the ratio image; step 5), fusing the filtered ratio image scene by scene with the filtered super image; step 6), performing enhanced Lee filtering on the N-scene fused image to obtain the N-scene final image, and completing multi-angle SAR image filtering. Although the above-mentioned existing methods have made certain progress in noise reduction, there is still a problem of insufficient detail retention.

[0004] In recent years, with the improvement of computing power and the development of machine learning technology, methods based on wavelet transform and methods based on sparse representation have been proposed. These methods improve denoising performance through processing in the transform domain. For example, the existing invention patent application document "SAR image denoising method based on swin transformer and wavelet transform" with publication number CN117036191A has the following implementation steps: performing wavelet decomposition of scale 1 on each sample in the generated noisy sample set to obtain 4 sub-band images of each sample, training the constructed Swin Transformer network, and denoising the SAR image. The present invention uses the SAR image denoising network of the swin transformer constructed by training samples after wavelet transform, and uses the multi-scale analysis characteristics of wavelet transform to perform local analysis on the signal at different scales, so that the edge texture information of the original image is retained while filtering out the high-frequency noise components; and uses the local window connection strategy of the swin transformer to enhance the image information characteristics while reducing the number of parameters of the model. However, these methods still face the problems of complex parameter selection, large amount of calculation, and insufficient processing of high-frequency information.

[0005] The emergence of deep learning technology has provided a new solution for SAR image denoising. Denoising methods based on deep learning, such as convolutional neural networks (CNNs), can achieve more effective denoising by learning deep features of large amounts of data. These methods can capture more complex image structures, but until recently, most deep learning methods have not been able to effectively handle speckle noise in SAR images while retaining important surface features.

[0006] There are several major drawbacks in existing technologies for denoising synthetic aperture radar (SAR) images:

[0007] Insufficient preservation of details and textures: Traditional denoising methods often use a uniform processing method, which causes the details of the denoised image to be smoothed together with the noise, thereby losing important surface features such as road edges, building outlines, and natural landforms.

[0008] Edge and structure blur: The edge and structure information in SAR images is crucial for target recognition and scene understanding. Existing technologies will weaken this key information while removing speckle noise, affecting the analysis and utilization of images.

[0009] Poor processing of high-frequency information: Since speckle noise mainly affects high-frequency areas, traditional methods tend to erase high-frequency information while suppressing noise, which is not conducive to small object detection and fine structure preservation.

[0010] Limited adaptability: Most traditional denoising techniques require manual selection of parameters, such as filter size and shape, which requires users to have a deep understanding of image content and noise characteristics, limiting the adaptability and application scope of the algorithm.

[0011] Weak generalization ability: Existing denoising algorithms are often optimized for specific types of noise or image conditions. When encountering new datasets or different noise patterns, the performance may drop sharply and lack generalization ability.

[0012] In summary, the existing technologies have technical problems such as insufficient detail and texture preservation, blurred edges and structures, poor high-frequency information processing, limited adaptability, and weak generalization ability. Summary of the invention

[0013] The technical problem to be solved by the present invention is: how to solve the technical problems of insufficient detail and texture preservation, blurred edges and structures, poor high-frequency information processing, limited adaptability and weak generalization ability in the prior art.

[0014] The present invention adopts the following technical solutions to solve the above technical problems: the SAR image denoising network method of adaptive multi-scale feature fusion AMFFD-Net includes:

[0015] S1. Collect and preprocess SAR images to generate noisy and noise-free image pairs to construct training sets and validation sets;

[0016] S2. Construct an adaptive multi-scale feature fusion model AMFFD-Net, wherein the adaptive multi-scale feature fusion model AMFFD-Net includes: a noise estimation module, an asymmetric convolution module, an adaptive feature selection module, a ReLU activation module, a multi-scale feature fusion module, a SE attention module, a cascade block and a deep residual learning module;

[0017] S3. Use the noise estimation module to estimate the SAR image noise to obtain denoising features; use the asymmetric convolution module to perform asymmetric convolution operations to extract convolution features from the SAR image; use the adaptive feature selection module to select correlation features based on the denoising features; use the ReLU activation module to apply the preset ReLU activation function to the correlation features, and process them to obtain multi-scale features; use the multi-scale feature fusion module to fuse multi-scale features to obtain fused features; use the SE attention module to use the preset attention mechanism to suppress irrelevant features in the fused features to obtain SAR fusion enhanced features; use the cascade block to perform convolution operations to learn the residual information in the SAR fusion enhanced features; use the deep residual learning module to perform deep residual learning based on the residual information to refine the SAR enhanced features to obtain SAR refined features; use the preset feature feedback loop to feed back the SAR refined features to the adaptive multi-scale feature fusion model AMFFD-Net, and perform cyclic refinement processing to obtain and output the denoised SAR image.

[0018] The synthetic aperture radar image denoising algorithm of the present invention significantly improves the denoising performance by introducing advanced technologies such as asymmetric convolution, adaptive feature selection, multi-scale feature fusion, and attention mechanism, and can maintain image details and edges while removing speckle noise. The combination of deep residual learning and feature feedback mechanism further enhances the learning ability and feature representation ability of the model.

[0019] The present invention develops and utilizes an adaptive denoising framework to automatically learn and adjust parameters to improve the generalization ability for different images and noise conditions. The present invention improves the adaptability and generalization ability, ensuring the robustness of the algorithm when dealing with different noise patterns and image contents. In addition, the optimized network structure and training strategy also improve the computational efficiency, making the algorithm suitable for large-scale and real-time SAR image processing needs, providing an effective and practical solution for SAR image denoising and quality improvement, and providing a fast and reliable image processing tool suitable for various application scenarios.

[0020] In a more specific technical solution, in step S3, the noise estimation module 1 is used to perform noise estimation on the SAR image according to the following logic:

[0021]

[0022] In the formula, is the estimated variance, I ij is the pixel value in the image, μ is the mean of the pixel values, and M and N are the dimensions of the image.

[0023] In a more specific technical solution, in step S3, the asymmetric convolution module 2 is used to perform an asymmetric convolution operation according to the following logic to capture high-frequency features from the SAR image:

[0024] I′=K*I

[0025] Where I′ is the convolved image, K is the convolution kernel, * represents the convolution operation, and I is the original image.

[0026] In a more specific technical solution, in step S3, the adaptive feature selection module AFS is used to dynamically adjust the processing strategy according to the image content, and the correlation features are selected according to the following logic to retain high-frequency information during the denoising process:

[0027] F selected =AFS(F all )

[0028] In the formula, F selected is the selected feature set, F all are all the extracted features, and AFS is the feature selection algorithm.

[0029] The present invention utilizes an asymmetric convolution module to improve the accuracy of processing the high-frequency part of an image, and utilizes an adaptive feature selection module to effectively identify and suppress speckle noise.

[0030] The asymmetric convolution module used in the present invention can more effectively capture high-frequency features in the image, such as edges and texture details, and maintain the clarity and details of the image during the denoising process. The adaptive feature selection module used in the present invention can ensure that high-frequency information is properly retained during the denoising process. By combining the asymmetric convolution module and the adaptive feature selection module, the present invention improves the network's ability to retain important image features while effectively suppressing noise.

[0031] In a more specific technical solution, in step S3, a ReLU activation module is used to apply the ReLU activation function to the correlation feature according to the following logic:

[0032] f(x)=max(0,x)

[0033] In the formula, x is the input feature, and f(x) is the feature after ReLU activation;

[0034] The following logic is used to concatenate the correlation features to obtain multi-scale features:

[0035] F concat =[F1; F2; ...; F n ]

[0036] In the formula, F concatis the concatenated feature vector, [;] represents the concatenation operation, F1, F2, ..., F b are characteristics of different layers.

[0037] In a more specific technical solution, in step S3, a multi-scale feature fusion module MFF is used to fuse multi-scale features according to the following logic:

[0038]

[0039] In the formula, F mff is the fused feature, w s is the weight of scale s, F s is the feature of scale s, and S is the set of all scales.

[0040] In a more specific technical solution, in step S3, the SE attention module is used to perform compression operation Squeeze, excitation operation Excitation and scaling operation Scale;

[0041] In the compression operation Squeeze, a global average pooling operation is performed using the following logic to compress the spatial dimension of the fused features:

[0042]

[0043] In the formula, z c is the compression feature of the cth channel, is a feature map, where H, W, and C are the height, width, and number of channels respectively. represents the output of global average pooling;

[0044] In the excitation operation Excitation, a fully connected layer FC is used to learn the dependencies between channels:

[0045] s=σ(g(z,W))=σ(W2δ(W1z))

[0046] In the formula, σ represents the Sigmoid activation function, δ represents the ReLU activation function, is the reduced-dimensional weight matrix, is the weight matrix of the increased dimension, r is the dimensionality reduction ratio, z is the compressed feature, Indicates incentive operation;

[0047] In the scaling operation Scale, the channels of the original feature map F are adjusted according to the output s of the excitation operation Excitation:

[0048]

[0049] In the formula, It is the feature map adjusted by SE block.

[0050] The present invention uses the SE attention module to perform compression, excitation, and scaling, and uses the attention mechanism to enhance important features while suppressing irrelevant features. The edge and structure clarity is enhanced, and by maintaining the structural integrity of the image, the edge and structure clarity of the denoised image is higher than that of the existing solution.

[0051] In a more specific technical solution, in step S3, the cascade block DCB7 is used to learn the residual information in the SAR fusion enhancement feature with the following logic:

[0052] F dcb =DCB(F input )

[0053] In the formula, F dcb is the feature output of the cascade block, F input is the input feature, and DCB represents the operation of the cascade block.

[0054] In a more specific technical solution, in step S3, a deep residual learning module is used to obtain refined features through deep residual learning according to the following logic:

[0055] F res =F input +Res(F input )

[0056] In the formula, F res is the feature after residual learning, F input is the input feature, and Res represents the residual mapping.

[0057] The present invention uses a denoising algorithm to perform denoising, which can effectively distinguish noise and detail information to retain more surface features and texture information. The present invention replaces the traditional uniform denoising method, retains the detail part and the noise part of the denoised image, and avoids the loss of surface features. The present invention is optimized for retaining high-frequency information to better handle small objects and fine structures. High-frequency information processing is optimized.

[0058] In a more specific technical solution, the SAR image denoising network system of adaptive multi-scale feature fusion AMFFD-Net includes:

[0059] The data set construction module is used to collect and preprocess SAR images to generate noisy and noise-free image pairs to construct training sets and validation sets;

[0060] A multi-scale feature model construction module is used to construct an adaptive multi-scale feature fusion model AMFFD-Net, wherein the adaptive multi-scale feature fusion model AMFFD-Net includes: a noise estimation module, an asymmetric convolution module, an adaptive feature selection module, a ReLU activation module, a multi-scale feature fusion module, a SE attention module, a cascade block and a deep residual learning module;

[0061] The SAR image noise estimation and removal module is used to estimate the SAR image noise using the noise estimation module to obtain denoising features; use the asymmetric convolution module to perform asymmetric convolution operations to extract convolution features from the SAR image; use the adaptive feature selection module to select correlation features based on the denoising features; use the ReLU activation module to apply the preset ReLU activation function to the correlation features to obtain multi-scale features based on processing; use the multi-scale feature fusion module to fuse the multi-scale features to obtain fusion features; use the SE attention module to use the preset attention mechanism to suppress irrelevant features in the fusion features to obtain SAR fusion enhancement features; use the cascade block to perform convolution operations to learn the residual information in the SAR fusion enhancement features; use the deep residual learning module to perform deep residual learning based on the residual information to refine the SAR enhancement features to obtain SAR refined features; use the preset feature feedback loop to feed back the SAR refined features to the adaptive multi-scale feature fusion model AMFFD-Net to perform refinement processing in a loop to obtain and output the denoised SAR image, and the SAR image noise estimation and removal module is connected to the multi-scale feature model construction module and the data set construction module.

[0062] Compared with the prior art, the present invention has the following advantages:

[0063] The synthetic aperture radar image denoising algorithm of the present invention significantly improves the denoising performance by introducing advanced technologies such as asymmetric convolution, adaptive feature selection, multi-scale feature fusion, and attention mechanism, and can maintain image details and edges while removing speckle noise. The combination of deep residual learning and feature feedback mechanism further enhances the learning ability and feature representation ability of the model.

[0064] The present invention develops and utilizes an adaptive denoising framework to automatically learn and adjust parameters to improve the generalization ability for different images and noise conditions. The present invention improves the adaptability and generalization ability, ensuring the robustness of the algorithm when dealing with different noise patterns and image contents. In addition, the optimized network structure and training strategy also improve the computational efficiency, making the algorithm suitable for large-scale and real-time SAR image processing needs, providing an effective and practical solution for SAR image denoising and quality improvement, and providing a fast and reliable image processing tool suitable for various application scenarios.

[0065] The present invention utilizes an asymmetric convolution module to improve the accuracy of processing the high-frequency part of an image, and utilizes an adaptive feature selection module to effectively identify and suppress speckle noise.

[0066] The asymmetric convolution module used in the present invention can more effectively capture high-frequency features in the image, such as edges and texture details, and maintain the clarity and details of the image during the denoising process. The adaptive feature selection module used in the present invention can ensure that high-frequency information is properly retained during the denoising process. By combining the asymmetric convolution module and the adaptive feature selection module, the present invention improves the network's ability to retain important image features while effectively suppressing noise.

[0067] The present invention uses the SE attention module to perform compression, excitation, and scaling, and uses the attention mechanism to enhance important features while suppressing irrelevant features. The edge and structure clarity is enhanced, and by maintaining the structural integrity of the image, the edge and structure clarity of the denoised image is higher than that of the existing solution.

[0068] The present invention uses a denoising algorithm to perform denoising, which can effectively distinguish noise and detail information to retain more surface features and texture information. The present invention replaces the traditional uniform denoising method, retains the detail part and the noise part of the denoised image, and avoids the loss of surface features. The present invention is optimized for retaining high-frequency information to better handle small objects and fine structures. High-frequency information processing is optimized.

[0069] The present invention solves the technical problems existing in the prior art of insufficient detail and texture preservation, blurred edges and structures, poor high-frequency information processing, limited adaptability and weak generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 A schematic diagram of the basic steps of the SAR image denoising network method of the adaptive multi-scale feature fusion AMFFD-Net according to Example 1 of the present invention;

[0071] Figure 2 This is a schematic diagram of data flow processing of the adaptive multi-scale feature fusion AMFFD-Net model according to Example 1 of the present invention;

[0072] Figure 3 This is a schematic diagram of the basic module connections of the adaptive multi-scale feature fusion AMFFD-Net model of Example 1 of the present invention. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] Example 1

[0075] like Figure 1 and Figure 2 As shown, the SAR image denoising network method of adaptive multi-scale feature fusion AMFFD-Net provided by the present invention includes the following basic steps:

[0076] S1. Data set preparation, generating training set and validation set;

[0077] In this embodiment, a preset number of SAR images are collected, and pairs of images with noise and images without noise are generated therefrom for training and verifying the model;

[0078] S2, data preprocessing;

[0079] In this embodiment, necessary preprocessing operations are performed on the collected SAR images. In this embodiment, the preprocessing operations include but are not limited to: image normalization and size adjustment;

[0080] S3. Build an adaptive multi-scale feature fusion model AMFFD-Net to process and output denoised SAR images.

[0081] like Figure 3 As shown, in this embodiment, the aforementioned adaptive multi-scale feature fusion model AMFFD-Net includes but is not limited to: a noise estimation module 1, an asymmetric convolution module 2, an adaptive feature selection module 3, a ReLU activation module 4, a multi-scale feature fusion module 5, a SE attention module 6, a cascade block 7 and a deep residual learning module 8;

[0082] In this embodiment, the asymmetric convolution module 2 is connected to the noise estimation module 1, the asymmetric convolution module 2 is connected to the adaptive feature selection module 3, the adaptive feature selection module 3 is connected to the ReLU activation module 4, the ReLU activation module 4 is connected to the multi-scale feature fusion module 5, the SE attention module 6 is connected to the multi-scale feature fusion module 5, the cascade block 7 is connected to the SE attention module 6, and the deep residual learning module 8 is connected to the cascade block 7; the deep residual learning module 8 is connected to the noise estimation module 1 through a feature feedback loop;

[0083] In the noise estimation module 1 of this embodiment, a noise estimation technique for analyzing an input SAR image is described using one processing box.

[0084]

[0085] in, is the estimated variance, I ij is the pixel value in the image, μ is the average pixel value, M and N are the image sizes;

[0086] In the asymmetric convolution module 2 of this embodiment, a processing box is used for asymmetric convolution operation to extract features from the image. This algorithm uses four types of convolutions: 3×1, 1×3, 3×3 and 5×5.

[0087] I′=K*I

[0088] Among them, I′ is the image after convolution, K is the convolution kernel, * represents the convolution operation, and I is the original image;

[0089] In the adaptive feature selection module 3 (AFS) of this embodiment, the following formula represents the adaptive feature selection module AFS selecting features based on the correlation of the denoising task:

[0090] F selected =AFS(F all )

[0091] Among them, F selected is the selected feature set, F all are all the extracted features, and AFS is the feature selection algorithm;

[0092] In this embodiment, an asymmetric convolution module and an adaptive feature selection module are used to retain high-frequency information. In this embodiment, an asymmetric convolution is used to more effectively capture high-frequency features in the image, and the high-frequency features include but are not limited to: edges and texture details, which are crucial for maintaining the clarity and details of the image during the denoising process;

[0093] In this embodiment, an adaptive feature selection module is used to dynamically adjust the processing strategy according to the image content to ensure that high-frequency information is properly retained during the denoising process. The combination of the asymmetric convolution module and the adaptive feature selection module improves the network's ability to retain important image features while effectively suppressing noise;

[0094] In this embodiment, an asymmetric convolution module is used to more accurately process the high-frequency part of the image, and an adaptive feature selection module is used to effectively identify and suppress speckle noise. The above design enables the network to distinguish important image details and noise during denoising, thereby reducing speckle noise while retaining key high-frequency information. In this embodiment, the synergy of the asymmetric convolution module and the adaptive feature selection module can improve the denoising effect, especially when processing SAR images with speckle noise;

[0095] In the ReLU activation module 4 of this embodiment, the ReLU activation function is applied to the selected features:

[0096] f(x)=max(0,x)

[0097] Among them, x is the input feature, f(x) is the feature after ReLU activation;

[0098] In this embodiment, feature splicing is performed to splice features of different layers or stages:

[0099] F concat =[F1; F2; ...; F n ]

[0100] Among them, F concat is the concatenated feature vector, [;] represents the concatenation operation, F1, F2, ..., F n are the characteristics of different layers;

[0101] In this embodiment, a multi-scale feature fusion module 5 (MFF) is used to fuse features of multiple scales:

[0102]

[0103] Among them, F mff is the fused feature, w s is the weight of scale s, F s is the characteristic of scale s, S is the set of all scales;

[0104] In this embodiment, the SE attention module 6 is used to enhance important features using the attention mechanism while suppressing irrelevant features. The basic operations of the SE block include: Squeeze, Excitation, and Scale.

[0105] In the compression operation Squeeze of this embodiment, global average pooling is used to compress the spatial dimension, and the feature map of each channel is compressed into a single value. If we have a feature map Where H, W, and C are the height, width, and number of channels respectively, and the output of the global average pooling Given by:

[0106]

[0107] Among them, z c is the compression feature of the cth channel;

[0108] In the excitation operation of this embodiment, a fully connected (FC) layer is used to learn the dependencies between channels. This step usually includes two FC layers, one for dimensionality reduction and the other for dimensionality increase. Usually, a ReLU activation function is used between these two FC layers. If z is a compressed feature, then the excitation operation It can be expressed as:

[0109] s=σ(g(z,W))=σ(W2δ(W1z))

[0110] Among them, σ represents the Sigmoid activation function, δ represents the ReLU activation function, is the reduced-dimensional weight matrix, is the weight matrix of dimension increase, r is the dimension reduction ratio;

[0111] In the scaling operation Scale of this embodiment, the output s of the excitation operation is used to adjust the channels of the original feature map F. This step is called scaling:

[0112]

[0113] in, It is the feature map adjusted by SE block;

[0114] In this embodiment, a series of convolution operations are represented by a cascade block DCB7 for learning residual information:

[0115] F dcb =DCB(F input )

[0116] Among them, F dcb is the feature output of the cascade block, F input is the input feature, DCB represents the operation of the cascade block;

[0117] In this embodiment, a deep residual learning module 8 is used to apply deep residual learning to refine features:

[0118] F res =F input +Res(F input )

[0119] Among them, F res is the feature after residual learning, F input is the input feature, Res represents the residual mapping;

[0120] In this embodiment, a feature feedback loop is used to feed back the refined features into the network for further refinement.

[0121] End the feedback loop and output the denoised SAR image.

[0122] The present invention is applicable to a variety of application scenarios, including but not limited to: remote sensing monitoring, military reconnaissance, aerospace applications, and millimeter wave imaging in security inspections. The present invention helps to improve the accuracy and efficiency of decision-making in these fields by improving image quality. This patent uses the MSTAR dataset and the millimeter wave dataset collected by the team.

[0123] In summary, the synthetic aperture radar image denoising algorithm of the present invention significantly improves the denoising performance by introducing advanced technologies such as asymmetric convolution, adaptive feature selection, multi-scale feature fusion, and attention mechanism, and can maintain image details and edges while removing speckle noise. The combination of deep residual learning and feature feedback mechanism further enhances the learning ability and feature representation ability of the model.

[0124] The present invention develops and utilizes an adaptive denoising framework to automatically learn and adjust parameters to improve the generalization ability for different images and noise conditions. The present invention improves the adaptability and generalization ability, ensuring the robustness of the algorithm when dealing with different noise patterns and image contents. In addition, the optimized network structure and training strategy also improve the computational efficiency, making the algorithm suitable for large-scale and real-time SAR image processing needs, providing an effective and practical solution for SAR image denoising and quality improvement, and providing a fast and reliable image processing tool suitable for various application scenarios.

[0125] The present invention utilizes an asymmetric convolution module to improve the accuracy of processing the high-frequency part of an image, and utilizes an adaptive feature selection module to effectively identify and suppress speckle noise.

[0126] The asymmetric convolution module used in the present invention can more effectively capture high-frequency features in the image, such as edges and texture details, and maintain the clarity and details of the image during the denoising process. The adaptive feature selection module used in the present invention can ensure that high-frequency information is properly retained during the denoising process. By combining the asymmetric convolution module and the adaptive feature selection module, the present invention improves the network's ability to retain important image features while effectively suppressing noise.

[0127] The present invention uses the SE attention module to perform compression, excitation, and scaling, and uses the attention mechanism to enhance important features while suppressing irrelevant features. The edge and structure clarity is enhanced, and by maintaining the structural integrity of the image, the edge and structure clarity of the denoised image is higher than that of the existing solution.

[0128] The present invention uses a denoising algorithm to perform denoising, which can effectively distinguish noise and detail information to retain more surface features and texture information. The present invention replaces the traditional uniform denoising method, retains the detail part and the noise part of the denoised image, and avoids the loss of surface features. The present invention is optimized for retaining high-frequency information to better handle small objects and fine structures. High-frequency information processing is optimized.

[0129] The present invention solves the technical problems existing in the prior art of insufficient detail and texture preservation, blurred edges and structures, poor high-frequency information processing, limited adaptability and weak generalization ability.

[0130] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. The SAR image denoising network method based on adaptive multi-scale feature fusion AMFFD-Net is characterized by: The method comprises: S1. Collect and preprocess SAR images to generate noisy and noise-free image pairs to construct training sets and validation sets; S2. Construct an adaptive multi-scale feature fusion model AMFFD-Net, wherein the adaptive multi-scale feature fusion model AMFFD-Net includes: a noise estimation module, an asymmetric convolution module, an adaptive feature selection module, a ReLU activation module, a multi-scale feature fusion module, a SE attention module, a cascade block and a deep residual learning module; S3, using the noise estimation module to estimate the SAR image noise to obtain denoising features; using the asymmetric convolution module to perform asymmetric convolution operations to extract convolution features from the SAR image; using the adaptive feature selection module to select correlation features according to the convolution features; using the ReLU activation module to apply the preset ReLU activation function to the correlation features, and thereby obtain multi-scale features; using the multi-scale feature fusion module to fuse the multi-scale features to obtain fusion features; using the SE attention module to suppress irrelevant features in the fusion features using the preset attention mechanism to obtain SAR fusion enhancement features; using the cascade block to perform convolution operations to learn the residual information in the SAR fusion enhancement features; using the deep residual learning module to perform deep residual learning according to the residual information to perform fine processing on the SAR fusion enhancement features to obtain SAR fine features; using a preset feature feedback loop to feed back the SAR fine features to the deep residual learning module to cyclically perform the fine processing to obtain and output a denoised SAR image.

2. The SAR image denoising network method of the adaptive multi-scale feature fusion AMFFD-Net according to claim 1 is characterized in that: In step S3, the noise estimation module 1 is used to perform noise estimation on the SAR image according to the following logic: In the formula, is the estimated variance, I ij is the pixel value in the image, μ is the mean of the pixel values, and M and N are the dimensions of the image.

3. The SAR image denoising network method of the adaptive multi-scale feature fusion AMFFD-Net according to claim 1 is characterized in that: In step S3, the asymmetric convolution module 2 is used to perform the asymmetric convolution operation according to the following logic to capture high-frequency features from the SAR image: I′=K*I Where I′ is the convolved image, K is the convolution kernel, * represents the convolution operation, and I is the original image.

4. The SAR image denoising network method of the adaptive multi-scale feature fusion AMFFD-Net according to claim 1 is characterized in that: In step S3, the adaptive feature selection module AFS is used to dynamically adjust the processing strategy according to the image content, and the correlation features are selected according to the following logic to retain high-frequency information during the denoising process: F selected =AFS(F all ) In the formula, F selected is the selected feature set, F all are all the extracted features, and AFS is the feature selection algorithm.

5. The SAR image denoising network method of the adaptive multi-scale feature fusion AMFFD-Net according to claim 1 is characterized in that: In step S3, the ReLU activation module is used to apply the ReLU activation function to the correlation feature according to the following logic: f(x) = max(0, x) In the formula, x is the input feature, and f(x) is the feature after ReLU activation; The correlation features are concatenated using the following logic to obtain the multi-scale features: F concat =[F1;F2;…;F n ] In the formula, F concat is the concatenated feature vector, [; ] represents the concatenation operation, F1, F2, ..., F n are characteristics of different layers.

6. The SAR image denoising network method of the adaptive multi-scale feature fusion AMFFD-Net according to claim 1 is characterized in that: In step S3, the multi-scale features are fused using the multi-scale feature fusion module MFF according to the following logic: In the formula, F mff is the fused feature, w s is the weight of scale s, F s is the feature of scale s, and S is the set of all scales.

7. The SAR image denoising network method of the adaptive multi-scale feature fusion AMFFD-Net according to claim 1 is characterized in that: In step S3, the SE attention module is used to perform a compression operation Squeeze, an excitation operation Excitation, and a scaling operation Scale; In the compression operation Squeeze, a global average pooling operation is performed using the following logic to compress the spatial dimension of the fused feature: In the formula, z c is the compression feature of the cth channel, is a feature map, where H, W, and C are the height, width, and number of channels, respectively. represents the output of global average pooling; In the excitation operation Excitation, a fully connected layer FC is used to learn the dependencies between channels: s=σ(g(z,W))=σ(W2δ(W1z)) In the formula, σ represents the Sigmoid activation function, δ represents the ReLU activation function, is the reduced-dimensional weight matrix, is the weight matrix of the increased dimension, r is the dimensionality reduction ratio, z is the compressed feature, Indicates incentive operation; In the scaling operation Scal e In the example, according to the output s of the excitation operation Excitation, the channels of the original feature map F are adjusted: In the formula, It is the feature map adjusted by SE block.

8. The SAR image denoising network method of the adaptive multi-scale feature fusion AMFFD-Net according to claim 1 is characterized in that: In step S3, the cascade block DCB7 is used to learn the residual information in the SAR fusion enhancement feature according to the following logic: F dcb =DCB(F input ) In the formula, F dcb is the feature output of the cascade block, F input is the input feature, and DCB represents the operation of the cascade block.

9. The SAR image denoising network method of adaptive multi-scale feature fusion AMFFD-Net according to claim 1 is characterized in that: In step S3, the deep residual learning module is used to obtain the refined features through deep residual learning according to the following logic: F res =F input +Res(F input ) In the formula, F res is the feature after residual learning, F input is the input feature, and Res represents the residual mapping.

10. The SAR image denoising network system based on adaptive multi-scale feature fusion AMFFD-Net is characterized by: The system comprises: The data set construction module is used to collect and preprocess SAR images to generate noisy and noise-free image pairs to construct training sets and validation sets; A multi-scale feature model construction module is used to construct an adaptive multi-scale feature fusion model AMFFD-Net, wherein the adaptive multi-scale feature fusion model AMFFD-Net includes: a noise estimation module, an asymmetric convolution module, an adaptive feature selection module, a ReLU activation module, a multi-scale feature fusion module, a SE attention module, a cascade block and a deep residual learning module; The SAR image noise estimation and removal module is used to estimate the SAR image noise using the noise estimation module to obtain denoising features; use the asymmetric convolution module to perform asymmetric convolution operations to extract convolution features from the SAR image; use the adaptive feature selection module to select correlation features according to the convolution features; use the ReLU activation module to apply the preset ReLU activation function to the correlation features to obtain multi-scale features; use the multi-scale feature fusion module to fuse the multi-scale features to obtain fused features; use the SE attention module to suppress the fused features using the preset attention mechanism The method comprises the steps of: extracting irrelevant features in the features to obtain SAR fusion enhanced features; using the cascade block to perform convolution operation to learn residual information in the SAR fusion enhanced features; using the deep residual learning module to perform deep residual learning according to the residual information to refine the SAR fusion enhanced features to obtain SAR refined features; using a preset feature feedback loop to feed back the SAR refined features to the deep residual learning module to cyclically perform the refined processing to obtain and output a denoised SAR image, and the SAR image noise estimation and removal module is connected to the multi-scale feature model construction module and the data set construction module.

Images