Seismic image mixed noise suppression method based on multi-attention feature adaptive fusion network
Through a multi-attention-based feature adaptive fusion network, the problems of long derivation time of existing seismic data denoising technology and limited application scenarios are solved, efficient denoising and detailed feature extraction of seismic images are achieved, and the denoising performance and generalization capabilities are improved.
Patent Information
- Application Number
- CN202411874326.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-19
- Publication Date
- 2025-03-28
AI Technical Summary
The existing seismic data denoising technology requires manual design and selection of features, resulting in a long model derivation time and modeling and assumptions based on specific noise types, limiting its application scenarios.
The seismic image mixed noise suppression method based on multi-attention feature adaptive fusion network is adopted. Through the dual-channel spatial attention module, the separation information extraction module and the comprehensive feature adaptive fusion module, the multi-scale feature information of the image is extracted and fused, and the degree of attention to different features is dynamically adjusted.
The feature extraction capability of the model for seismic image details is improved, the denoising performance is enhanced, and a variety of mixed noise can be suppressed more effectively, which improves the quality of the denoising image and the generalization ability of the model.
Smart Images
Figure SMS_1 
Figure SMS_2 
Figure QLYQS_1
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer technology, in particular to the field of data enhancement and image denoising technology, and relates to a method for denoising multiple types of noise images, in particular to a method for suppressing mixed noise of seismic images based on a multi-attention feature adaptive fusion network. Background Art
[0002] Traditional image denoising models are divided into four categories: spatial domain, transform domain, natural statistics, and sparse representation. The spatial domain-based method directly filters the pixels of the image. For example, the median filter reduces noise by calculating the median of the neighborhood around the pixel. The method is simple and computationally efficient, but it may cause problems such as blurred image details and boundary fusion. The transform domain-based method converts the image to the frequency domain or wavelet domain for processing. For example, BLS-GSM models and reconstructs the wavelet coefficients to remove noise, but it is not effective in dealing with complex noise scenes. The natural statistics-based method uses the natural statistical characteristics of the image to denoise. For example, BM3D estimates and suppresses noise based on the similarity between image blocks, but it can only handle specific types of noise. The sparse representation method assumes that the image has a sparse representation in a certain domain. For example, NLSC decomposes the image into a sparse representation form and then removes noise through sparse coding, but it has high computational complexity and high training cost.
[0003] The denoising method based on deep learning uses deep neural networks to learn the mapping relationship between noisy images and clean images, and accurately reconstructs the details and structure of noisy images by training the network, minimizing the difference between the real denoised image and the network output image to achieve the purpose of image denoising. Among them, generative adversarial networks, convolutional neural networks, and encoders are widely used in image denoising. Based on the generative adversarial network CGAN, the generator learns a mapping function to transform the noisy image into a matching clean image, and the discriminator learns a classifier that can accurately judge the authenticity of the image and distinguish the generated denoised image from the real clean image, but the training process is unstable and produces artifacts and over-smoothing. Based on the convolutional neural network PA-MRNet, parallel multi-resolution convolution flow is used to extract multi-resolution features, parallel attention module captures important features, and parallel attention guides the fusion module to fuse the extracted features, which can restore weak signals with less capacity loss. Based on the encoder, the Swin Transformer performs self-attention calculations in the sliding window, saves computing resources and can extract long-range information dependencies of the image. It can accurately reconstruct seismic signals and suppress noise in seismic data at the same time, but the training time is long.
[0004] At present, most seismic data denoising technologies require manual design and feature selection, which leads to long model derivation time and are based on modeling and assumptions of specific noise types, thus limiting their application scenarios. Therefore, it is urgent to design a denoising solution that can accurately reconstruct seismic data, effectively suppress mixed noise (random noise, multiple waves) and have generalization capabilities, improve the model's ability to extract features of detailed information, and enhance denoising performance. Summary of the invention
[0005] 1. Purpose of the Invention
[0006] The purpose of the present invention is to provide a seismic image mixed noise suppression method based on a multi-attention feature adaptive fusion network for seismic images containing multiple types of noise, so that the model can better perceive the subtle texture information of the seismic image, so as to better restore the shallow texture features of the seismic image, thereby achieving the purpose of improving the quality of the denoised image.
[0007] 2. Technical solution
[0008] 1. A method for suppressing mixed noise in seismic images based on a multi-attention feature adaptive fusion network, characterized by comprising the following steps:
[0009] Step 1: Use m seismic images of the actual work area with original size of S1×S2 containing unknown random noise and multiple waves to divide the training set and the test set according to a certain ratio;
[0010] Where m∈Z + ; S1,S2∈Z + ;
[0011] Step 2: Define the range containing the complete multiple waves in the training set image, and replace the rest with a 0 mask. Perform the same operation on the noisy image and the clean image to ensure that the image size remains unchanged, and obtain n data-enhanced images of the training set;
[0012] Where n∈Z + ;
[0013] Step 3: Construct an information fusion module based on multi-attention, separation information extraction and comprehensive feature adaptive fusion, which includes the following structure:
[0014] Step 3.1: Construct a dual-channel spatial attention module to extract key channel information and spatial information in the image. The module includes the following structure:
[0015] First, the channel attention mechanism and spatial attention mechanism are used to process the image to obtain the channel feature F of the image. C and spatial feature F S , concatenate the two in the channel dimension to get F CS; Next, the layer normalization eliminates the distribution difference of each layer input, and then uses M1×M2 convolution to extract feature information F E , after the GELU activation function and then with F E Multiply, strengthen the important feature information, suppress the unimportant feature information, realize the function of feature filtering, and obtain the feature information F G ; Then, the channel attention mechanism and spatial attention mechanism are used to process F respectively G , and obtain the secondary channel features F' C and the quadratic space characteristic F' S Finally, concatenate in the channel dimension to obtain the comprehensive output feature F O1 ;
[0016] Among them, the channel attention mechanism obtains global features through global average pooling, maps global features through 1×1 convolution layer, PReLU activation function and 1×1 convolution, and then uses Sigmoid activation function to map to probability, multiply with input features at pixel level, and assign different attention weights to each channel; the spatial attention mechanism uses global average pooling and global maximum pooling respectively, splices them in the channel dimension, combines the overall background information and important information, maps global features through 5×5 convolution, maps to probability using Sigmoid activation function, multiplies with input features at pixel level, and assigns attention weights to different spatial positions;
[0017] Step 3.2: Construct a separation information extraction module to fuse multi-channel feature information. The module includes the following structure:
[0018] First, rectangular convolutions of sizes 1×3, 3×5, and 5×7 are used to extract features of different scales, and F 13 、F 35 、F 57 , concatenate in the channel dimension, use 3×3 convolution with dilation=6, increase the effective receptive field of the convolution kernel to focus on a wider range of context information, and get F d6 ; Then, three features F1, F2, and F3 are separated in the channel dimension, and information of different scales in different channel feature maps is extracted through 3×1, 5×3, and 7×5 rectangular convolutions respectively. Then, they are concatenated in the channel dimension and 3×3 convolution with dilation=6 is used to obtain F' d6 ; Then, F d6 and F' d6 Add them together to get F" d6 , concatenate F in the channel dimension d6 , F' d6 and F"d6 Get F d2 ; Finally, F d6 After global average pooling and extraction of global features, d2 Multiply pixel by pixel to get F' d2 , use 1×1 convolution to integrate the features between different channels and get F O2 ;
[0019] Among them, F d6 =F d6 +F d6 ';F d2 '=GAP(F d6 )F d2 ;
[0020] Step 3.3: Construct a comprehensive feature adaptive fusion module to fuse feature information of different scales and weight it to feature information of each scale, and then integrate the weighted feature information. The module includes the following structure:
[0021] First, the three input information X1, X2, and X3 are added to obtain X a , X1, X2, X3 and X a The feature information X that integrates multi-scale information is obtained by splicing in the channel dimension c Then, the global features are obtained through global average pooling, mapped through 1×1 convolution, GELU activation function and 3×3 convolution, and then mapped to probability P through Softmax activation function. s ; Finally, the input information X1, X2 and X3 are respectively s Multiply and weight the original input information to get X g1 , X g2 and X g3 , X g1 , X g2 and X g3 Add together to get X g , X is converted into g1 , X g2 , X g3 and X g Splice to get X gc , use 1×1 convolution to reduce the dimension of feature information and get X O ;
[0022] Among them, X a =X1+X2+X3, X c =concate(X1,X2,X3,X a ), X g =X g1 +X g2 +Xg3 , X gc =concate(X g1 ,X g2 ,X g3 ,X g ), Concate is a concatenation operation; Softmax function:
[0023] Step 4: Construct a hybrid loss function to train the multi-attention feature adaptive fusion network. The calculation formula is as follows:
[0024] Network training loss function: Loss = MSE + α × Perception_loss, including mean square error MSE focusing on pixel-level differences and perceptual loss Perception_loss focusing on high-level features of the image through pre-trained VGG16;
[0025] in, N is the number of pixels, y i is the pixel value of the label image, is the pixel value of the denoised image; Perception_loss is calculated by sending the denoised image and the label image to the pre-trained VGG16 network using MSE;
[0026] Step 5: Introduce the improved feature extraction module constructed in step 3 into the loss function of step 4 and design and combine it into a complete multi-attention feature adaptive fusion network for training:
[0027] Step 5.1: Construct a multi-attention feature adaptive fusion network. This module includes the following structure:
[0028] First, the input feature information X undergoes a 1×1 convolution X 1C , processed through three channels of different scales. The first channel: extract the key channel information and spatial information in the image through the dual-channel spatial attention module to obtain F O1 After and X 1C Add them together to get F1; the second channel: pass through a downsampling module to get X 2C , after the separation information extraction module fuses the multi-channel feature information, we get F O2 After and X 2C Add, and perform an upsampling module and a dropout layer with rate = N1 to get F2; the third channel: after two consecutive downsampling modules, get X 3C, after the self-attention mechanism and two consecutive upsampling modules, F3 is obtained; then, F1, F2 and F3 are fused with feature information of different scales through the comprehensive feature adaptive fusion module and weighted to feature information of each scale, and then the weighted feature information is integrated and passed through the dropout layer with rate = N2 to obtain F t ; Finally, F1, F2, F3 and F t Add them together to get F xt , F1, F2, F3 and F xt Then concatenate in the channel dimension and obtain F after 1×1 convolution dimensionality reduction. O ;
[0029] Among them, downsampling module; Concate is a concatenation operation, C = Concate (C 13 ,C 23 ,C 33 );
[0030] Step 5.2: Set the learning rate Lr, Lr∈(0,0.001); batch size Bz, Bz∈Z + ; Training period E, E∈Z + ; and optimizer Adam;
[0031] Step 5.3: Put the seismic noise images of the training set S1×S2 and their corresponding pure images into the model for training, and save the model after a complete training cycle;
[0032] Step 6: Call the model after a complete training cycle for testing: put the test set earthquake noise images of size S1×S2 into the model to obtain the denoising results, and obtain the earthquake denoising images of size S1×S2.
[0033] The innovative features of the present invention are as follows:
[0034] (1) Compared with the traditional CBAM module, the dual-channel spatial attention module adopts two channel attention and spatial attention mechanisms to form two sets of parallel channel spatial attention modules. The two sets of parallel channel spatial attention modules are cascaded and gated convolution is added between them to filter out non-important feature information and capture more important channel and spatial feature information.
[0035] (2) Compared with the traditional information extraction module, the above-mentioned separation information extraction module adopts dilated convolution to fuse the features extracted by the multi-channel convolution structure, divides the feature information in the channel dimension, and then uses multi-channel convolution to extract features separately. It also uses global average pooling to obtain the global feature information and then multiplies it with the feature map of the feature fusion. It automatically adjusts the weight of each feature according to the global information, so that the model can dynamically adjust the degree of attention to different features under different input images.
[0036] (3) Compared with the traditional feature fusion module, the comprehensive feature adaptive fusion module adds the three inputs and then concatenates them with the three inputs in the channel dimension to obtain comprehensive feature information. It uses GAP to extract global features, multiplies the comprehensive features with the three inputs respectively, optimizes feature selection, and integrates different feature information.
[0037] 3. Beneficial effects
[0038] The present invention discloses a method for suppressing noise of seismic images, which is a method for suppressing noise of multiple types of mixed noise images. A method for suppressing mixed noise of seismic images based on a multi-attention feature adaptive fusion network is provided. Firstly, important channel and spatial position information of the seismic image is obtained through a dual-channel spatial attention module, and important structural information of the noisy image is reconstructed. Feature information of different scales is separated through a separation information extraction module, and texture information of the image is analyzed. Finally, a comprehensive feature adaptive fusion module is used to fuse richer features and retain more original subtle feature information. This method not only ensures the effective noise suppression purpose of the model, but also improves the image restoration accuracy and generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 This is the algorithm flow chart of the feature adaptive fusion network based on multi-attention. First, different multiple waves, different levels of Gaussian noise, salt and pepper noise, and multiplicative noise are added to the clean seismic image. Then, the noisy image and the clean seismic image are sent to the denoising network for denoising, and the denoised image is output. The loss function is used to calculate the difference between the denoised image and the clean seismic image, and the denoising network is guided by back propagation to suppress noise.
[0040] Figure 2 This is the structure diagram of the dual-channel spatial attention module. The input data size is unified as 360×160. It first passes through the channel and spatial attention mechanisms at the same time, and then performs feature fusion on the channel splicing. Then it passes through the gated convolution to filter out the unimportant feature information. Then, it passes through the channel and spatial attention mechanisms at the same time to obtain the important position information of the feature map. Finally, it passes through the channel dimension splicing feature fusion to form a dual-channel spatial attention structure.
[0041] Figure 3 To separate the information extraction module structure diagram, first, the input 180×80 data is subjected to a three-channel convolution operation, and the features of the 3×3 convolution with dilation=6 are concatenated in the channel dimension to obtain the intermediate features. Figure 1 ; For intermediate features Figure 1The channel dimension is split into intermediate feature maps 11, 12, and 13. The three feature maps are respectively subjected to three-channel convolution operations, spliced in the channel dimension, and fused by 3×3 convolution with dilation=6 to obtain the intermediate feature. Figure 2 ; The intermediate features Figure 1 and intermediate features Figure 2 Add together to get the intermediate features Figure 3 ; Then, the intermediate features Figure 1 , intermediate features Figure 2 and intermediate features Figure 3 Perform feature fusion on the channel to obtain intermediate features Figure 4 ; The intermediate features Figure 1 The global features obtained by global average pooling and the intermediate features Figure 4 After multiplication and 1×1 convolution for feature fusion, the output features are obtained;
[0042] Figure 4 The structure diagram of the multi-channel information fusion module is as follows: The three input data X1, X2, and X3 of 360×160 are added to obtain X a , X1, X2, X3 and X in the channel dimension a Splice and get the comprehensive feature X c ; After global average pooling, the global features are obtained, and the channel weight β is obtained after convolution layer mapping; β is multiplied with X1, X2, and X3 respectively to obtain X a1 , X a2 , X a3 ; X a1 , X a2 , X a3 Add together to get X at , X is converted into a1 , X a2 , X a3 , X at Splicing is performed, and feature fusion is performed through 1×1 convolution to obtain output features;
[0043] Figure 5 The structure of the complete multi-attention feature adaptive fusion network;
[0044] Figure 6 The effect of denoising using the multi-attention feature adaptive fusion network. The first column is the real sub-image after artificial denoising, the second column is the sub-image containing random noise and multiple waves, and the third column is the model denoising; DETAILED DESCRIPTION
[0045] In order to make the purpose, technical scheme and beneficial effects of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, and the embodiments use partial fault data of oil fields. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention, that is, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments.
[0046] Step 1: Use 1260 seismic images of the actual work area with an original size of 360×160 containing unknown random noise and multiple waves to divide them into 1200 training sets and 60 test sets according to the proportion;
[0047] Step 2: In the training set images, the range containing the complete multiple waves is delineated, and the rest is replaced by a 0 mask. The same operation is performed on the noisy image and the clean image to ensure that the image size remains unchanged, and 712 data-enhanced images of the training set are obtained;
[0048] Step 3: Construct an information fusion module based on multi-attention, separation information extraction and comprehensive feature adaptive fusion. The module includes the following structure:
[0049] Step 3.1: Construct a dual-channel spatial attention module to extract key channel information and spatial information in the image;
[0050] Step 3.2: Construct a separation information extraction module to fuse multi-channel feature information;
[0051] Step 3.3: Construct a comprehensive feature adaptive fusion module to fuse feature information of different scales and weight it to feature information of each scale, and then integrate the weighted feature information;
[0052] Step 4: Construct a hybrid loss function to train a multi-attention feature adaptive fusion network;
[0053] Step 5: Introduce the improved feature extraction module constructed in step 3 into the loss function of step 4 and design and combine it into a complete multi-attention feature adaptive fusion network for training:
[0054] Step 5.1: Construct a multi-attention feature adaptive fusion network. This module includes the following structure:
[0055] First, the input feature information X undergoes a 1×1 convolution X 1C , processed through three channels of different scales. The first channel: extract the key channel information and spatial information in the image through the dual-channel spatial attention module to obtain F O1 After and X 1C Add them together to get F1; the second channel: pass through a downsampling module to get X 2C, after the separation information extraction module fuses the multi-channel feature information, we get F O2 After and X 2C Add, and perform an upsampling module and a dropout layer with rate = N1 to get F2; the third channel: after two consecutive downsampling modules, get X 3C , after the self-attention mechanism and two consecutive upsampling modules, F3 is obtained; then, F1, F2 and F3 are fused with feature information of different scales through the comprehensive feature adaptive fusion module and weighted to feature information of each scale, and then the weighted feature information is integrated and passed through the dropout layer with rate = N2 to obtain F t ; Finally, F1, F2, F3 and F t Add them together to get F xt , F1, F2, F3 and F xt After concatenation in the channel dimension and dimensionality reduction by 1×1 convolution, we get F O ;
[0056] Among them, downsampling module; Concate is a concatenation operation, C = Concate (C 13 ,C 23 ,C 33 );
[0057] Step 5.2: Set the learning rate Lr, Lr∈(0,0.001); batch size Bz, Bz∈Z + ; Training period E, E∈Z + ; and optimizer Adam;
[0058] Step 5.3: Put the 360×160 training set seismic noise images and their corresponding pure images into the model for training, and save the model after a complete training cycle;
[0059] Step 6: Call the model after a complete training cycle for testing: Put the test set earthquake noise image with a size of 360×160 into the model to obtain the denoising result, and obtain a 360×160 earthquake denoising image.
[0060] The following table compares the denoising effect between this method and different methods. The evaluation index is the peak signal-to-noise ratio PSNR, and the calculation formula is: Unit: decibel (dB); Structural similarity SSIM, the calculation formula is Where x and y represent the original image and the denoised image respectively, μ x and μ y Represent the pixel mean of the two images, and represents the variance of the two images, σ xyRepresents the covariance of the two images; c1 and c2 are constants greater than zero to avoid the denominator being zero;
[0061] Table 1 Comparison of recognition accuracy between the method of the present invention and different models
[0062] Tab.1 Our method improves the recognition rate of various models
[0063] Model PSNR SSIM WACAFRN 27.80 0.8569 Swin Transformer 28.31 0.8596 PA-MRNet 29.31 0.8774 3_SIDN 29.47 0.8758 Algorithm of the present invention 31.64 0.9109
Claims
1. A method for suppressing mixed noise in seismic images based on a multi-attention feature adaptive fusion network, characterized in that The following steps are involved: Step 1: Use m seismic images of the actual work area with original size of S1×S2 containing unknown random noise and multiple waves to divide the training set and the test set according to a certain ratio; Where m∈Z + ; S1,S2∈Z + ; Step 2: In the training set image, define the range containing the complete multiple waves, and replace the rest with a 0 mask. Perform the same operation on the noisy image and the clean image to ensure that the image size remains unchanged, and obtain n data-enhanced images of the training set; Where n∈Z + ; Step 3: Construct an information fusion module based on multi-attention, separation information extraction and comprehensive feature adaptive fusion. The module includes the following structure: Step 3.1: Construct a dual-channel spatial attention module to extract key channel information and spatial information in the image. The module includes the following structure: First, the channel attention mechanism and spatial attention mechanism are used to process the image to obtain the channel feature F of the image. C and spatial feature F S , concatenate the two in the channel dimension to get F CS ; Next, the layer normalization eliminates the distribution difference of each layer input, and then uses M1×M1 convolution to extract feature information F E , after the GELU activation function and then with F E Multiply, strengthen the important feature information, suppress the unimportant feature information, realize the function of feature filtering, and obtain the feature information F G ; Then, the channel attention mechanism and spatial attention mechanism are used to process F respectively G , and obtain the secondary channel features F' C and the quadratic space characteristic F' S Finally, concatenate in the channel dimension to obtain the comprehensive output feature F O1 ; Among them, the channel attention mechanism obtains global features through global average pooling, maps global features through 1×1 convolution layer, PReLU activation function and 1×1 convolution, and then uses Sigmoid activation function to map to probability, multiply with input features at pixel level, and assign different attention weights to each channel; the spatial attention mechanism uses global average pooling and global maximum pooling respectively, splices them in the channel dimension, combines the overall background information and important information, maps global features through 5×5 convolution, maps to probability using Sigmoid activation function, multiplies with input features at pixel level, and assigns attention weights to different spatial positions; Step 3.2: Construct a separation information extraction module to fuse multi-channel feature information. The module includes the following structure: First, rectangular convolutions of sizes 1×3, 3×5, and 5×7 are used to extract features of different scales, and F 13 、F 35 、F 57 , concatenate in the channel dimension, use 3×3 convolution with dilation=6, increase the effective receptive field of the convolution kernel to focus on a wider range of context information, and get F d6 ; Then, three features F1, F2, and F3 are separated in the channel dimension, and information of different scales in different channel feature maps is extracted through 3×1, 5×3, and 7×5 rectangular convolutions respectively. Then, they are concatenated in the channel dimension and 3×3 convolution with dilation=6 is used to obtain F' d6 ; Then, F d6 and F' d6 Add them together to get F" d6 , concatenate F in the channel dimension d6 、F' d6 and F" d6 , and get F d2 ; Finally, F d6 After global average pooling and extraction of global features, d2 Multiply pixel by pixel to get F' d2 , use 1×1 convolution to integrate the features between different channels and get F O2 ; Among them, F d6 =F d6 +F d6 ';F d2 '=GAP(F d6 )F d2 ; Step 3.3: Construct a comprehensive feature adaptive fusion module to fuse feature information of different scales and weight it to feature information of each scale, and then integrate the weighted feature information. The module includes the following structure: First, the three input information X1, X2, and X3 are added to obtain X a , X1, X2, X3 and X a The feature information X that integrates multi-scale information is obtained by splicing in the channel dimension c Then, the global features are obtained through global average pooling, mapped through 1×1 convolution, GELU activation function and 3×3 convolution, and then mapped to probability P through Softmax activation function. s ; Finally, the input information X1, X2 and X3 are respectively s Multiply and weight the original input information to get X g1 , X g2 and X g3 , X g1 , X g2 and X g3 Add together to get X g , X is converted into g1 , X g2 , X g3 and X g Splice to get X gc , use 1×1 convolution to reduce the dimension of feature information and get X O ; Among them, X a =X1+X2+X3, X c =concate(X1,X2,X3,X a ), X g =X g1 +X g2 +X g3 , X gc =concate(X g1 ,X g2 ,X g3 ,X g ), Concate is a concatenation operation; Softmax function: Step 4: Construct a hybrid loss function to train the multi-attention feature adaptive fusion network. The calculation formula is as follows: Network training loss function: Loss = MSE + α × Perception_loss, including mean square error MSE focusing on pixel-level differences and perceptual loss Perception_loss focusing on high-level features of the image through pre-trained VGG16; in, N is the number of pixels, y i is the pixel value of the label image, is the pixel value of the denoised image; Perception_loss is calculated by sending the denoised image and the label image to the pre-trained VGG16 network using MSE; Step 5: Introduce the improved feature extraction module constructed in step 3 into the loss function of step 4 and design and combine it into a complete multi-attention feature adaptive fusion network for training: Step 5.1: Construct a multi-attention feature adaptive fusion network. This module includes the following structure: First, the input feature information X undergoes a 1×1 convolution X 1C , processed through three channels of different scales. The first channel: extract the key channel information and spatial information in the image through the dual-channel spatial attention module to obtain F O1 After and X 1C Add them together to get F1; the second channel: pass through a downsampling module to get X 2C , after the separation information extraction module fuses the multi-channel feature information, we get F O2 After and X 2C Add, and perform an upsampling module and a dropout layer with rate = N1 to get F2; the third channel: after two consecutive downsampling modules, get X 3C , after the self-attention mechanism and two consecutive upsampling modules, F3 is obtained; then, F1, F2 and F3 are fused with feature information of different scales through the comprehensive feature adaptive fusion module and weighted to feature information of each scale, and then the weighted feature information is integrated and passed through the dropout layer with rate = N2 to obtain F t ; Finally, F1, F2, F3 and F t Add them together to get F xt , F1, F2, F3 and F xt After concatenation in the channel dimension and dimensionality reduction by 1×1 convolution, we get F O ; Among them, downsampling module; Concate is a concatenation operation, C = Concate (C 13 ,C 23 ,C 33 ); Step 5.2: Set the learning rate Lr, Lr∈(0,0.001); batch size Bz, Bz∈Z + ; Training period E, E∈Z + ; and optimizer Adam; Step 5.3: Put the seismic noise images of the training set S1×S2 and their corresponding pure images into the model for training, and save the model after a complete training cycle; Step 6: Call the model after a complete training cycle for testing: put the test set earthquake noise images of size S1×S2 into the model to obtain the denoising results, and obtain the earthquake denoising images of size S1×S2.