A seismic data random noise blind denoising method based on gradient penalty generative adversarial network

Abstract

The application provides a seismic data random noise blind denoising method based on a generative adversarial network. First, the original data set is proportionally divided into training data and test data, and Gaussian noise and unknown random noise are added in the training and test data respectively; during training, the data is sent into an improved residual module for first coarse feature extraction, so as to ensure the correlation between adjacent image pixels; then, an improved channel attention mechanism is used for second fine extraction of the feature vector, so as to ensure the completeness of the image itself; finally, the improved gradient penalty function is calculated to enlarge the gap between the real samples and the false samples, so as to not only save the training time, but also enhance the discrimination ability of the discriminator and improve the feature extraction ability of the generator. The network has good generalization and high accuracy, and has good performance in the denoising of some complex unknown noise seismic images.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of computer, especially the technical field of data enhancement and image denoising, and relates to a method for realizing blind denoising of images containing different random noises, in particular to an image denoising method based on an improved gradient penalty generative adversarial network. BACKGROUND

[0002] There are four main categories of traditional image denoising models, namely, spatial domain-based, transform domain-based, natural statistics-based and sparse representation-based. Among them, the typical method is the median filter method, which is a relatively popular spatial domain-based method. However, this method ignores the characteristics of each pixel in the image and tends to grasp the overall characteristics, resulting in phenomena such as blurred details and boundary fusion in the denoised image. The BLS-GSM based on the transform domain can preserve some characteristics of the pixels, but the overall effect is poor. The NLSC based on sparse representation can preserve more characteristic information, but the calculation time is longer and the training cost is high. The BM3D based on natural statistics can only remove a certain type of noise.

[0003] Based on the deep learning denoising method, the deep image denoising model is proposed by using the nonlinearity, high efficiency and depth of deep learning. Among them, the self-encoding network, convolutional neural network and generative adversarial network are widely used in image denoising field. The DCAENN based on self-encoding network can better remove the known noise during training, but it cannot perform blind denoising for different noises or unknown noises, and has poor generalization ability. The stack sparse denoising self-encoding network combines multiple stacks to combine the sparse coding and denoising coding of the pre-trained deep neural network for image denoising. However, this method is a supervised algorithm and can only denoise the known noise of the training set, and has very limited blind denoising ability. The ResNet based on residual network can process Gaussian noise with unknown noise level, but cannot remove different types of noise. The basic conditional generative adversarial network GAN repeatedly discriminates noise data and label data, which achieves the purpose of blind denoising, but the training difficulty is large, and the gradient vanishing or gradient explosion may occur.

[0004] At present, most of the seismic data denoising technologies have slow model parameter derivation speed, difficult detail effect and edge feature discrimination due to their inherent heavy computation, thereby limiting their application in practical scenarios. Therefore, it is urgent to propose a denoising scheme with high precision and strong generalization ability, so as to improve the noise image perception ability of the model and enhance the denoising performance. SUMMARY

[0005] 1. OBJECTIVE

[0006] The purpose of the present application is to provide a seismic data random noise blind denoising method based on an improved gradient penalty generative adversarial network, so that the model can better fit the real feature information of the picture, not limited to a single noise, better express the structural characteristics of the picture, and achieve the purpose of blind denoising, aiming at the noise contained in the seismic image extracted from the actual work area at present is the superposition of multiple random noises, not a single Gaussian noise.

[0007] 2. Technical solution

[0008] A seismic data random noise blind denoising method based on a gradient penalty generative adversarial network, characterized by comprising the following steps:

[0009] Step 1: Use m original seismic data pictures with unknown noise containing an original size of S1xS2 to make a training set and a test set, first, divide the m section pictures into a training set and a test set according to a certain proportion, and perform preprocessing operation on the size of each picture, and process them into a size of (S1-n)x(S2-m), then cut each preprocessed picture, and cut each picture into k sub-pictures with a size of 100x100, and the total number of sub-pictures is M, wherein the total number of sub-pictures in the training set is and the total number of sub-pictures in the test set is , forming the training set and the test set required for model training;

[0010] Wherein, m∈Z + ; S1,S2∈Z + , S1,S2>100; (S1-n) and (S2-n) are integer multiples of 100; k∈Z + ; the total number of sub-pictures M=mk; Downward rounding is represented; λ∈(0,1);

[0011] Step 2: Construct a feature extraction module based on the fusion of improved residual and improved channel attention mechanism, and the improved feature fusion module comprises the following structure:

[0012] Step 2.1: Construct an improved residual feature extraction module to coarsely extract data features, and the module comprises the following structure:

[0013] First, given a batch of data x i , then calculate the mean value μ i of the batch of data, and then perform mean normalization, that is, subtract the mean value, to obtain Next, calculate the variance σ 2 of the batch of data, and then perform variance normalization, that is, divide by the variance, and introduce a small constant ε to prevent the divisor from being zero, to obtain The normalized samples are then scaled and shifted, and the distribution is adjusted using learnable parameters γ and β to obtain the final normalized output feature y. ij Using activation function σ L For y ij Activation is performed to obtain the activated distribution X′, ensuring the non-zero nature of X′. Finally, X′ is processed by an adaptive convolution block with kernel size K to concatenate the extracted information H with the original information in the channel dimension, forming a new fusion feature H′.

[0014] Where, x i ∈Z and satisfy mean of batch data x ij This represents the pixel value in the i-th row and j-th column, where m ∈ Z. + Mean normalization Batch data variance Variance normalization Output features LeakyRelu functions: For y ij Activate all data: X′=σ L (Bn(s1,s2...s n The formula for adaptive convolution kernel size is: C represents the number of feature channels, K represents the size of the adaptively selected convolutional kernel; Concate is the concatenation operation, H′=Concate(X,H);

[0015] Step 2.2: Construct an improved channel attention mechanism-based module to further refine the information H′ after feature fusion. This feature extraction module includes the following structure:

[0016] First, the input fused feature information H′ is subjected to global average pooling in the x and y directions, respectively, to obtain two feature vectors of 1×100 and 100×1. Then, these two feature vectors are fused by matrix multiplication to obtain the fused feature map H″. Next, this feature map is input into a feature extraction module with a three-channel multi-scale dilated convolution for multi-scale feature stacking. The stacked feature information is concatenated along the channels using Concate, resulting in feature map Z. Finally, the stacked feature map Z is input into a convolution module with a kernel size of 1×1 for dimensionality reduction and σ... S The activation function is used to activate the feature, resulting in new feature weights {w1, w2, ..., w}. k Finally, the fused feature information H′ output from step 3 is concatenated on the channel to obtain the fused feature information H″′ with the improved attention mechanism.

[0017] The formula for calculating the input feature information H′ to H″ is as follows: and This indicates that average pooling is performed in the x and y directions respectively; the three-channel multi-scale dilated convolutional feature extraction module has convolutional kernel sizes of 3×3, 5×5, and 7×7 with dilation rates of 1, 2, and 3 respectively; Concate is the concatenation of channels; Z = Concate(H1″, H2″, H3″), where H1″, H2″, and H3″ represent the feature information extracted through the three channels respectively; the weight calculation formula is: w i =σ S (CID k (Z)), σ S (·) is the Sigmoid activation function, CID k This represents a two-dimensional convolution operation with an adaptive convolution kernel;

[0018] Step 3: Introduce a new gradient penalty term to calculate the gradient of the discriminator and increase the feature difference between real and fake samples. The formula is as follows:

[0019] First, the Generative Adversarial Network (WGAN) with a distance function is introduced to calculate the feature distance, |f(x1)-f(x2)|≤k|x1-x2|. Then, L is treated as the target to calculate the loss, so the loss of the generator G can be simplified as G(Loss)=-D(Loss), and the loss of the discriminator D is D(Loss)=D(X). i Next, a new gradient penalty function is introduced to widen the gap between adjacent features: Finally, calculate the loss error under the new gradient penalty function: grad p en = MSE(Norm-k), and the new loss value: L = D(real) - (D(X) i ))+10×(α+β)×grad p en;

[0020] Where f(x1) is the true quantity, f(x2) is the quantity with bias, and k represents the distribution coefficient of the bias over distance, k∈R + G(Loss) is the generator loss value, D(Loss) is the discriminator loss value, and D(X) is the generator loss value. i ) is the discriminant value of the i-th noisy data, D(real i ) represents the label of the i-th real data; Norm represents the calculated new gradient, grad represents the calculated gradient; α and β are the dynamic expansion factors for dynamic gradient expansion. D(X i-1D(real) represents the discriminant value of the previous noisy data. i-1 ) represents the label of the previous manually denoised data; grad p en is the loss value error under the new gradient penalty; MSE is the root mean square error.

[0021] Step 4: Introduce the improved feature extraction module constructed in Step 2 into the new gradient penalty function in Step 3 and design and combine them into a complete generator network G. At the same time, design a suitable discriminator network D for adversarial training.

[0022] Step 4.1: First, construct an A and B dual-channel generator G network, with each channel including four feature extraction modules that fuse the improved residual and improved channel attention mechanism described in Step 2;

[0023] Step 4.2: Construct a discriminator consisting of n convolutional structures for adversarial learning, where n∈Z. + ;

[0024] Step 4.3: Set the learning rate Lr, Lr∈(0,0.1); batch size Bz, Bz∈Z + Training period E, E∈Z + ; and the optimizer Adam;

[0025] Step 4.4: Transform the actual work area A 100×100 image of earthquake noise, cut out and its corresponding manually denoised image, is fed into the model for training. The model is then saved after a complete training cycle.

[0026] Step 5: Test the model after a complete training cycle: An earthquake noise image with a size of 100×100 was put into the model for denoising test, and all the denoised sub-images were restored to the original size of S1×S2 according to the cutting direction and size.

[0027] 3. The innovative aspects of this invention are as follows:

[0028] (1) Compared with traditional residual networks, the above-mentioned improved residual module adopts a combination of activation and convolution, which ensures the non-zero nature of the input convolutional layer data. At the same time, it adopts adaptive convolution instead of fixed convolution operation, which improves the number of features and recognition accuracy while maintaining the network complexity.

[0029] (2) Compared with the traditional channel attention mechanism, the above-mentioned improved channel attention mechanism uses matrix multiplication to achieve feature fusion when processing average pooling in different directions, instead of simple channel splicing. It also uses three-channel multi-scale dilated convolutional blocks to extract features, which improves the feature difference between adjacent pixels and also improves the recognition accuracy.

[0030] (3) Compared with the generative adversarial network WGAN with distance function, the gradient penalty loss function is improved, which can further expand the distance features calculated by the traditional loss function, greatly improve the model's perception and improve the internal correlation between adjacent features.

[0031] 4. Beneficial effects

[0032] This invention discloses a blind denoising method for seismic data based on gradient-penalized generative adversarial networks, addressing the issue that noise in actual seismic images from real-world work areas is often random noise resulting from the superposition of multiple noise sources rather than single Gaussian noise. First, an improved residual network structure reduces most of the noise in the seismic image. Then, an improved channel attention mechanism module adjusts the details of the image, improving the model's structural similarity. Finally, an improved gradient-penalized loss function further corrects the gap between the real and generated image structures. This method achieves the goal of blind denoising while also improving the model's accuracy and generalization ability. Attached Figure Description

[0033] Figure 1 To generate the overall algorithm flowchart of the adversarial network, the original dataset is input into the generator to generate images. The generated images and real data are then input into the discriminator to determine their authenticity. Data that is judged to be real is saved into the model parameters and used for the next training. Data that is judged to be fake is returned to the starting position to regenerate and train.

[0034] Figure 2 The module structure diagram for improving the residual structure of generator G is shown. The input data size is uniformly 100×100. After batch normalization, it enters the LeakyRelu activation function for activation, and then enters the adaptive convolutional layer for feature extraction. The above operation is repeated twice, and the output and input are concatenated in the channel to form the residual structure.

[0035] Figure 3The module structure diagram for improving the channel attention mechanism of generator G is as follows: First, the input 100×100 data is subjected to global average pooling in the x and y directions to obtain 1×100 and 100×1 feature vectors, which are then multiplied by matrix. Then, a three-channel multi-scale dilated convolution operation is performed with convolution kernel sizes of 3×3, 5×5, and 7×7 and dilation rates of 1, 2, and 3, respectively, to extract features and concatenate them in the channels. Finally, after a 3×3 convolution operation, a two-channel convolution dimensionality reduction operation is performed with a 1×1 convolution kernel and a Sigmoid activation function.

[0036] Figure 4 For the complete structure diagram of the improved gradient-penalized generative adversarial network generator G, the improved residual module and channel attention mechanism module are spliced ​​into a whole module P. There are four such modules P on each channel, for a total of two channels.

[0037] Figure 5 The first column is the real sub-image after denoising using gradient penalty to generate an adversarial network, the second column is the sub-image with random noise, and the third column is the sub-image after denoising using the model.

[0038] Figure 6 To create an overall image by denoising all the effect sub-images and then stitching them together, the left side is the real image after manual denoising, the middle is the image with random noise, and the right side is the effect image after model denoising. Detailed Implementation

[0039] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. The embodiments use data from 30 self-made simulated gun sets. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of this invention, and not all of them.

[0040] Step 1: Use 30 original 301×1501 seismic data images containing unknown noise from the actual work area to create training and test sets. First, divide the 30 cross-sectional images into training and test sets according to a certain ratio. Preprocess the size of each image to 300×1500. Then, cut each preprocessed image into 45 sub-images with a length and width of 100×100, resulting in a total of 1350 sub-images. Among them, there are 1080 training sub-images and 270 test sub-images, forming the training and test sets required for model training.

[0041] Step 2: Construct a feature extraction module based on the fusion of improved residual and improved channel attention mechanisms. The improved feature fusion module includes the following structure:

[0042] Step 2.1: Construct an improved residual feature extraction module to coarsely extract data features;

[0043] Step 2.2: Construct an improved channel attention mechanism-based module to further refine the information H′ after the above-mentioned fused features;

[0044] Step 3: Introduce a new gradient penalty term to calculate the gradient and widen the feature gap between real and fake samples in the discriminator;

[0045] Step 4: Introduce the improved feature extraction module constructed in Step 2 into the new gradient penalty function in Step 3 and design and combine them into a complete generator network G. At the same time, design a suitable discriminator network D for adversarial training.

[0046] Step 4.1: First, construct an A and B dual-channel generator G network, with each channel including four feature extraction modules that fuse the improved residual and improved channel attention mechanism described in Step 2;

[0047] Step 4.2: Construct a discriminator composed of n convolutional structures for learning adversarial skills, where n∈Z+;

[0048] Step 4.3: Set the learning rate Lr = 0.0001, batch size Bz = 10, training epochs E = 20000, and optimizer Adam;

[0049] Step 4.4: Input 1080 100×100 size earthquake noise images cut from the actual work area and their corresponding manually denoised images into the model for training, and save the model after a complete training cycle.

[0050] Step 5: Test the model after a complete training cycle: Put 270 earthquake noise images of size 100×100 into the model for denoising test, and restore all the denoised sub-images to the original size of 300×1500 according to the cutting direction and size.

[0051] Table 1 compares the denoising performance of this method with different methods, where the evaluation metric is the peak signal-to-noise ratio (PSNR), calculated using the following formula: Unit: decibel (dB); Structural similarity score (SSIM), calculated using the following formula: Where x and y represent the original image and the denoised image, respectively, and μ x and μ y These represent the pixel mean values ​​of the two images, respectively. and σ represents the variance of two images. xy This represents the covariance of the two graphs; c1 and c2 are constants greater than zero, used to avoid zero denominators.

[0052] Table 1 Comparison of recognition accuracy between the method of the present invention and different models

[0053] Tab.1 Our method improves the recognition rate of various models

[0054] Model PSNR SSIM WGAN-GP 24.25 0.66 GR-GAN 25.4 0.69 ResNet18 26.68 0.77 Algorithm of the present invention 29.21 0.82

Claims

1. A blind denoising method for random noise in seismic data based on generative adversarial networks, characterized in that... Includes the following steps: Step 1: Using m original seismic images of size S1×S2 containing unknown noise from the actual work area, create training and test sets. First, divide the m cross-sectional images into training and test sets according to a certain ratio. Preprocess the size of each image to (S1-n)×(S2-m). Then, cut each preprocessed image into k sub-images with dimensions of 100×100. A total of M sub-images are cut from the m cross-sectional images, of which the training set sub-images contain a total of... Zhang, test set graph Zhang, to form the training set and test set needed for model training; Where, m∈Z + S1, S2 ∈ Z + S1, S2 > 100; (S1-n) and (S2-n) are both integer multiples of 100; k ∈ Z + The total number of subgraphs is M = mk; This indicates rounding down; λ∈(0,1); Step 2: Construct a feature extraction module based on the fusion of improved residual and improved channel attention mechanisms. The improved feature fusion module includes the following structure: Step 2.1: Construct an improved residual feature extraction module to perform coarse extraction of data features. This module includes the following structure: First, given a batch of data x i Then calculate the mean μ of this batch of data. j Then perform mean normalization, that is, subtract the mean, to get Next, calculate the variance σ of this batch of data. 2 Then, variance normalization is performed, i.e., dividing by the variance, and a small constant ε is introduced to prevent the divisor from being zero, resulting in... Then, the normalized samples are scaled and shifted, and the distribution is adjusted using learnable parameters γ and β to obtain the final normalized output feature y. ij Using activation function σ L For y ij Activation is performed to obtain the activated distribution X′, ensuring the non-zero nature of X′. Finally, X′ is processed by an adaptive convolution block with kernel size K to concatenate the extracted information H with the original information in the channel dimension, forming a new fusion feature H′. Where, x i ∈Z and satisfy mean of batch data x ij This represents the pixel value in the i-th row and j-th column, where m ∈ Z. + Mean normalization Batch data variance m∈Z + Variance normalization ε→0 + Output features γ,β∈(0,1); LeakyReLU function: For y ij Activate all data; adaptive convolution kernel size: C represents the number of feature channels, K represents the size of the adaptively selected convolutional kernel; Concate is the concatenation operation, H′=Concate(X,H); Step 2.2: Construct an improved channel attention mechanism-based module to further refine the information H′ after feature fusion. This feature extraction module includes the following structure: First, the input fused feature information H′ is subjected to global average pooling in the x and y directions, respectively, to obtain two feature vectors of 1×100 and 100×1. Then, these two feature vectors are fused by matrix multiplication to obtain the fused feature map H″. Next, this feature map is input into a feature extraction module with a three-channel multi-scale dilated convolution for multi-scale feature stacking. The stacked feature information is concatenated along the channels using Concate, resulting in feature map Z. Finally, the stacked feature map Z is input into a convolution module with a kernel size of 1×1 for dimensionality reduction and σ... S The activation function is used to activate the feature, resulting in new feature weights w. i Finally, the fusion feature information H′ output from step 2.1 is concatenated on the channel to obtain the fusion feature information H″′ after the improved attention mechanism. The formula for calculating the input feature information H′ to H″ is as follows: and This indicates that average pooling is performed in the x and y directions respectively; the three-channel multi-scale dilated convolutional feature extraction module has convolutional kernel sizes of 3×3, 5×5, and 7×7 with dilation rates of 1, 2, and 3 respectively; Concate is the concatenation of channels; Z = Concate(H1″, H2″, H3″), where H1″, H2″, and H3″ represent the feature information extracted through the three channels respectively; the weight calculation formula is: w i =σ S (CID k (Z)), σ S () is the Sigmoid activation function, CID k This represents a two-dimensional convolution operation with an adaptive convolution kernel; Step 3: Introduce a new gradient penalty term to calculate the gradient of the discriminator and increase the feature difference between real and fake samples. The formula is as follows: First, the Generative Adversarial Network (WGAN) with a distance function is introduced to calculate the feature distance, |f(x1)-f(x2)|≤k|x1-x2|. Then, L is treated as the target to calculate the loss, so the loss of the generator G can be simplified as G(Loss)=-D(Loss), and the loss of the discriminator D is D(Loss)=D(X). i Next, a new gradient penalty function is introduced to widen the gap between adjacent features: Finally, calculate the loss error under the new gradient penalty function: grad p en = MSE(Norm-k), and the new loss value: L = D(real) - (D(X) i ))+10×(α+β)×grad p en; Where f(x1) is the true quantity, f(x2) is the quantity with bias, and k represents the distribution coefficient of the bias over distance, k∈R + G(Loss) is the generator loss value, D(Loss) is the discriminator loss value, and D(X) is the generator loss value. i ) is the discriminant value of the i-th noisy data, D(real i ) represents the label of the i-th real data; Norm represents the calculated new gradient, grad represents the calculated gradient; α and β are the dynamic expansion factors for dynamic gradient expansion. D(X i-1 D(real) represents the discriminant value of the previous noisy data. i-1 ) represents the label of the previous manually denoised data; grad p en is the loss value error under the new gradient penalty; MSE is the root mean square error. Step 4: Introduce the improved feature extraction module constructed in Step 2 into the new gradient penalty function in Step 3 and design and combine them into a complete generator network G. At the same time, design a suitable discriminator network D for adversarial training. Step 4.1: First, construct an A and B dual-channel generator G network, with each channel including four feature extraction modules that fuse the improved residual and improved channel attention mechanism described in Step 2; Step 4.2: Construct a discriminator consisting of n convolutional structures for adversarial learning, where n∈Z. + ; Step 4.3: Set the learning rate Lr, Lr∈(0,0.1); batch size Bz, Bz∈Z + Training period E, E∈Z + ; and the optimizer Adam; Step 4.4: Transform the actual work area The 100×100 size earthquake noise sub-images that have been cut and their corresponding manually denoised sub-images are put into the model for training, and the model after a complete training cycle is saved. Step 5: Test the model after a complete training cycle: An earthquake noise sub-image with a size of 100×100 was put into the model for denoising test, and all the denoised sub-images were restored to the original size of S1×S2 according to the cutting direction and size.

Images