Automatic denoising method based on non-matching geological images

Abstract

Current geological sedimentary models contain a large amount of noise, which is randomly distributed and difficult to remove manually. Existing geological model image denoising methods have high requirements for dataset preparation and complex denoising steps. Therefore, we propose an automatic denoising method based on mismatched geological images. We use multiple noisy images obtained through a cropping algorithm and a few noise-free images, cutting them at different intervals to construct a mismatched noise-noise dataset that does not correspond to river channels. Subsequently, based on style transfer, we treat the noisy images as the first type of images and the noise-free images as the second type of images, and propose an automatic denoising method based on generative adversarial networks. This method aims to output the corresponding denoising result for the input noisy image. This method simplifies the steps of existing denoising methods by eliminating the need for noise detection before denoising. Denoising results for various river channels show that this method can effectively remove incorrect noise, thus ensuring the integrity of the river channels.

Description

Technical Field

[0001] This invention pertains to methods for geological denoising and computer vision style transfer, and relates to the fields of computer vision and geological exploration. Background Technology

[0002] In the field of geology, geological modeling has always played a crucial role. Over the past decade, numerous scholars have made significant contributions to geological modeling methods. Geological modeling techniques have evolved from two-point geostatistical modeling to multi-point geostatistical modeling. Two-point geostatistical modeling utilizes the Kriging method to establish deterministic models and applies stochastic simulation methods to establish optional, equally probable geological models. Later, many scholars researched multi-point geostatistical modeling. The Snesim algorithm uses training images instead of the variogram between two points to reproduce reservoir configurations with complex structures. The Simpat algorithm incorporates similarity calculations to compare the similarity between data events at the point to be estimated and training images, thereby enhancing the selectivity of geological models. Filtersim uses filters for pattern classification, utilizing overall patterns to reduce the computational burden. However, these modeling techniques are still insufficient to provide reasonably descriptive 3D geological models. Geostatistical modeling is moving towards an intelligent modeling stage. Specifically, the improved Generative Adversarial Network (WGAN) is used to train models, directly generating realistic geological models with a small number of random parameters. However, deep learning-based modeling methods require a large number of training samples. While deep learning-based prediction networks can improve the quality of geological data, they cannot prevent noise caused by data errors. Current work focuses on denoising geological data. For example, self-supervised deep learning networks are applied to seismic data denoising. It is worth noting that noise still appears when printing models. Therefore, utilizing deep learning networks to remove noise from sedimentary model images is crucial.

[0003] In the field of computer vision, geological image denoising can be viewed as a process of finding outliers. Because noise is difficult to label, general supervised anomaly detection networks are unsuitable for geological image denoising. Therefore, reconstruction-based anomaly detection methods have attracted considerable attention. Firstly, several anomaly detection algorithms based on convolutional autoencoder (AE) networks have emerged. For example, structured energy-based AEs have been used to model training samples; researchers have proposed jointly encoding vectors and reconstruction errors in AEs. It is worth noting that the performance of AEs is affected by insufficient latent space design. To address this issue, generative adversarial networks (GANs) have been used in anomaly detection. GANs can simulate complex high-dimensional distributions, especially images, through a mini-maximum game process, and the model can successfully generate data with excellent performance. Schlegl et al. proposed AnoGAN, which uses a convolutional structure similar to DCGAN to build a generator trained only on normal images, and then calculates anomaly scores based on the reconstructed image and the discriminator's feature maps to discover anomaly markers in medical images. To reduce the complexity of the mapping from image to latent space, Zenati et al. effectively utilized the BiGAN structure to distinguish anomalies with the same anomaly scores. In subsequent research, Akcay et al. proposed GANomaly, which includes an encoder-decoder-encoder network group to explore deep latent representations of normal samples and employs anomaly scores computed from the latent space.

[0004] Currently, based on the aforementioned anomaly detection methods, noise detection and removal methods for geological model images have been proposed. For example, unsupervised geological image denoising methods train the generator on noise-free images, enabling the generator to learn the distribution of noise-free features and thus perform a restoration function. Subsequently, noise-based detection and removal methods for datasets with noise and noise-free images have also been proposed, which make it easier for the model to capture noise locations and river discontinuities by comparing noisy and noise-free images. However, these methods require first detecting noise in the geological images and then repeatedly denoising based on the detection results, making the denoising process complex and varied. In addition, the paired dataset approach requires a high degree of correspondence between the river channels in the two images, and the dataset creation is relatively complex. Therefore, we propose an automatic denoising method for non-matched geological images that does not require pairing or noise detection and is based on the idea of ​​style transfer. Summary of the Invention

[0005] The purpose of this invention is to improve the original denoising method, which has high requirements for the correspondence between river channels in the dataset and has complicated steps for denoising geological images. Therefore, an automatic denoising method based on training on a non-matching dataset is proposed.

[0006] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0007] S1. Implement an automatic cutting algorithm for geological images with noise based on existing geological models.

[0008] S2. Based on S1, multiple noisy geological images and noise-free images are segmented at intervals to obtain a non-matching noisy-noise image dataset.

[0009] S3. Construct an automatic denoising model based on non-matching geological images.

[0010] S4. Train the denoising model constructed in S3 based on the dataset in S2.

[0011] Most denoising methods rely on manually cropping the original image, resulting in noisy geological images often featuring black borders, and the manual cropping process is tedious. Therefore, automatically converting 3D geological models into geological images using a computer is a much more convenient method. First, geological software is used to convert the 3D geological model into a [0,1] code in "csv" format. Here, 0 represents blue non-river areas, and 1 represents red river channels. This file is then divided into multiple "csv" files of size 774×546. Given a 774×546 all-black image, color correction is performed on the black image based on the [0,1] code in one of the obtained csv files. Specifically, we iterate through the codes in the csv file, with rows of 774 and columns of 546. If the code is "1", the corresponding position in the black image is changed to white (255, 255, 255); if the code is "0", the pixel value at the corresponding position in the black image is not modified. Following this process, we obtain a binarized noisy image. Then, the binarized image is filled with color, resulting in a noisy geological image with red and blue contrasts. This process can be represented by the following formula:

[0012] (z1,z2,z3……)=cut_data(Z csv (1)

[0013] (img1,img2,img3...)=change(F(img black ,(z1,z2,z3))) (2)

[0014] Among them, Z csv The "csv" file represents the model transformation, where z1, z2, and z3 represent the transformations from Z. csv The resulting "csv" files contain the segmented images. `cut_data()` represents the file segmentation process. (img) blackThe first image represents a completely black image, and the second, img1, img2, and img3 represent the segmented geological images with noise. F represents the pixel modification process, and change() represents the color change process.

[0015] The geological images obtained by this method are free from the interference of black borders, avoiding the preprocessing of noisy images during manual cropping, thus simplifying the image acquisition and processing process.

[0016] Current methods for denoising geological sedimentary model images primarily use datasets consisting of noise-free images or noisy-noise image pairs, requiring a one-to-one correspondence between the river channel shapes in the noisy and noise-free images. Therefore, datasets constructed by other methods place high demands on the correspondence between noise-free and noisy images. To address this issue, this invention constructs a non-matching dataset of noisy and noise-free images without any corresponding relationships.

[0017] To ensure that there is no direct correspondence between noisy and noise-free images in the constructed dataset, we simultaneously segmented multiple noisy geological images of different shapes with a noise-free image of a different shape. We segmented the noisy geological images to 128×128 pixels at intervals of 16, 32, and 64 pixels, obtaining a total of 780 noisy images. We then segmented the noise-free image at intervals and flipped it to obtain 780 noise-free images. This invention uses a mismatched noisy-noise image dataset to eliminate the one-to-one correspondence between paired datasets.

[0018] Current unsupervised methods for denoising geological sedimentary model images first utilize a reconstruction-based generative adversarial network (GAN) to generate a reconstructed image from an input noise-free image, minimizing the distance between the two images to allow the generator to learn the morphological distribution of the noise-free image. However, these methods have limited denoising effectiveness. Subsequently, supervised geological image denoising methods capture noise and discontinuities in rivers by comparing the differences between the input noisy and noise-free images. However, these methods have high requirements for the dataset, ensuring a one-to-one correspondence between the noisy and noise-free river channels in each input image. Furthermore, both the unsupervised and supervised denoising methods mentioned above involve two steps: first, noise detection is performed on the geological image, followed by denoising, making the denoising process complex and cumbersome. Therefore, this invention proposes an automatic denoising method based on non-matching geological images. This method eliminates the need for dataset pairing preprocessing, and its generated result is directly the denoised geological image, greatly simplifying the denoising process and making it more convenient and easy to use.

[0019] This invention's automatic denoising method based on non-matching geological images utilizes the concept of style transfer, classifying noisy geological images as Class A and noise-free images as Class B. In deep learning, style transfer involves inputting a "brown horse" image into a model to generate a "striped horse" image, without altering the horse's shape and size. This type of model is well-suited for handling noisy images input into a model to obtain their corresponding noise-free images, while preserving the shape and color of the river channel.

[0020] The automatic denoising model comprises two generators with identical network structures and two discriminators with identical structures. The generators are built upon a convolutional autoencoder (DAE), and the discriminators are structured similarly to an encoder. Since the primary objective of this invention is to input a noisy image into the generator to obtain its corresponding noise-free image, noise is added to the input noisy image to prevent Gaussian noise-like interference from affecting the quality of the generated image. This allows the generator to learn a more useful feature distribution.

[0021] We first input a noisy image X. noise Noise is added to it to obtain geological images with additional noise. The image is then input into the generator G, which is used to generate a noise-free image. A2B This results in a generated image that resembles a noise-free image. Then, we X... good Input to a generator G for generating noisy images. B2A This results in a generated image that resembles a noisy image. To enhance the generation effect of the two generators, we will generate images. Then input into generator G B2A This generates a looping image that resembles a noisy image. Then generate the image Then input into generator G A2B This results in a loop image that resembles a noise-free image. This process can be represented by the following formula:

[0022]

[0023]

[0024]

[0025]

[0026] Here, `add_noise()` represents adding additional noise to the original noisy input image. θ g1 ,θg2 These are the generator G A2B and generator G B2A Intermediate process parameters.

[0027] Then we will generate the image Compared with the original geological noise image X noise Input to discriminator D A We will compare the generated image with the original image. With noise-free image X good Input to discriminator D B Perform the original / generated distinction. This process can be represented by the following formula:

[0028]

[0029]

[0030] in, Discriminator D A The feature vectors output from the last layer correspond to X respectively. noise , Discriminator D B The feature vectors output from the last layer correspond to X respectively. good ,

[0031] The generator training for an automatic denoising method based on mismatched geological images mainly consists of three parts. First, we take the noise-free image X... good Input to generator G A2B Thus, the generated image is obtained. We will X-ray the noisy geological images noise Input to generator G B2A Obtain the generated image We minimize With image X good Distance between and minimization With image X noise The distance between them. This process can be represented as:

[0032]

[0033]

[0034] Where, loss ide_B With loss ide_A Representing the generator G respectively A2B With generator G B2A Image contrast loss.

[0035] The model uses a max-min game approach to train the generator and discriminator. The discriminator is fixed during generator training. For the generator G... A2B It expects to generate images Discriminator D B The result is true; for generator G B2A It expects to generate images Discriminator D A The result is determined to be true. This process can be represented as:

[0036]

[0037]

[0038] Where, loss gan_A2B With loss gan_B2A These are the generator G A2B With generator G B2A The loss of competitive training.

[0039] Furthermore, we minimize the loop image when training the generator. The distance between it and its corresponding original image. This process can be represented as:

[0040]

[0041]

[0042] Where, loss cyc_B With loss cyc_A It is the minimum contrast loss between the looped image from the two generators and the original image.

[0043] In summary, we can obtain the total loss of the generator. G It can be expressed by the following formula:

[0044] Loss G =α×(loss) ide_B +loss ide_A )+β×(loss gan_A2B +loss gan_B2A )+ρ×(loss cyc_B +loss cyc_A (15)

[0045] Among them, α=5, β=1, and ρ=10. Loss G This is the sum of the six losses mentioned above.

[0046] When training the discriminator, we fix the generator and first train the discriminator D. A Its purpose is to accurately identify the real image X.noise With generated image Then the discriminator D was trained. B Its purpose is to accurately identify the real image X. good With generated image Their training loss A With loss B It can be expressed by the following formula:

[0047]

[0048]

[0049] We implement this method using PyTorch. First, we redesign the encoder, decoder, and discriminator based on a convolutional neural network. During training, we load images of size (batch, 3, 64, 64) at a time, where the batch size is set to 5. The output image size is (batch, 3, 64, 64), and the discriminator's output vector dimension is (batch, 1). Additionally, we set the learning rate to 2×10n. -4 The Adam optimizer is used for gradient updates.

[0050] Compared with existing technologies, the beneficial effects of this invention are:

[0051] 1. This invention proposes an automatic denoising method based on generative adversarial networks, based on the idea of ​​style transfer. Unlike other denoising methods, the generated result of this model is the denoised result, thus avoiding the noise detection step and simplifying the denoising operation of geological images.

[0052] 2. This invention obtains a non-matching noise-noise image dataset by segmenting multiple existing geological noise images and a few noise-free images at intervals. This not only reduces the need for river channel correspondence in other denoising methods, but also avoids the trouble of manual labeling. Attached Figure Description

[0053] Figure 1 Flowchart for creating a noisy-noise-free image dataset with no matching

[0054] Figure 2 This is a model structure diagram of an automatic denoising method based on mismatched geological images.

[0055] Figure 3 A small geological image with noise and its corresponding denoised image.

[0056] Figure 4 Original noisy geological image and its corresponding denoising result

[0057] Figure 5 Comparison of the results of generating a model once and generating a model twice for the same geological image.

[0058] Figure 6 The results of the unsupervised denoising method are compared with the results of this patent and the similarity between them and the original noise-free image. Detailed Implementation

[0059] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0060] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0061] Figure 1 This is a flowchart illustrating the creation of a noisy-to-noise image dataset with no matching. Most denoising methods rely on manually cropping the original image, resulting in noisy geological images often having black borders, and the manual cropping process is tedious. Therefore, automatically converting 3D geological models into geological images via computer is a convenient method. First, geological software is used to convert the 3D geological model into a [0,1] code in "csv" format. Here, 0 represents blue non-river areas, and 1 represents red river channels. This file is then divided into multiple "csv" files of size 774×546. Given a 774×546 all-black image, color correction is performed on the black image based on the [0,1] code in one of the obtained csv files. Specifically, we iterate through the codes in the csv file, with rows of 774 and columns of 546. If the code is "1", the corresponding position in the black image is modified to white (255, 255, 255); if the code is "0", the pixel value at the corresponding position in the black image is not modified. Following this process, we obtain a binarized noisy image. Then, we fill the binarized image with color, thus obtaining a noisy geological image with red and blue contrasts. This process can be expressed by the following formula:

[0062] (z1,z2,z3……)=cut_data(Z csv (1)

[0063] (img1,img2,img3...)=change(F(img black ,(z1,z2,z3))) (2)

[0064] Among them, Z csv The "csv" file represents the model transformation, where z1, z2, and z3 represent the transformations from Z. csv The resulting "csv" files contain the segmented images. `cut_data()` represents the file segmentation process. (img) blackThe first image represents a completely black image, and the second, img1, img2, and img3 represent the segmented geological images with noise. F represents the pixel modification process, and change() represents the color change process.

[0065] The geological images obtained by this method are free from the interference of black borders, avoiding the preprocessing of noisy images during manual cropping, thus simplifying the image acquisition and processing process.

[0066] Current methods for denoising geological sedimentary model images primarily use datasets consisting of noise-free images or noisy-noise image pairs, requiring a one-to-one correspondence between the river channel shapes in the noisy and noise-free images. Therefore, datasets constructed by other methods place high demands on the correspondence between noise-free and noisy images. To address this issue, this invention constructs a non-matching dataset of noisy and noise-free images without any corresponding relationships.

[0067] To ensure that there is no direct correspondence between noisy and noise-free images in the constructed dataset, we simultaneously segmented multiple noisy geological images of different shapes with a noise-free image of a different shape. We segmented the noisy geological images to 128×128 pixels at intervals of 16, 32, and 64 pixels, obtaining a total of 780 noisy images. We then segmented the noise-free image at intervals and flipped it to obtain 780 noise-free images. This invention uses a mismatched noisy-noise image dataset to eliminate the one-to-one correspondence between paired datasets.

[0068] Figure 2 This diagram illustrates the model structure of an automatic denoising method based on non-matching geological images. Current unsupervised geological sedimentary model image denoising methods first utilize a reconstruction-based generative adversarial network (GAN) to generate a reconstructed image from an input noise-free image, minimizing the distance between the two images to allow the generator to learn the morphological distribution of the noise-free image. However, these methods have limited denoising effectiveness. Subsequently, supervised geological image denoising methods capture noise and discontinuities in rivers by comparing the differences between the input noisy and noise-free images. However, these methods require a high-quality dataset, ensuring a one-to-one correspondence between the noisy and noise-free river channels in each input image. Furthermore, both the unsupervised and supervised denoising methods involve two steps: noise detection followed by denoising, making the denoising process complex and cumbersome. Therefore, this invention proposes an automatic denoising method based on non-matching geological images. This method eliminates the need for dataset pairing preprocessing, and its generated result is directly the denoised geological image, greatly simplifying the denoising process and making it more convenient and user-friendly.

[0069] This invention's automatic denoising method based on non-matching geological images utilizes the concept of style transfer, classifying noisy geological images as Class A and noise-free images as Class B. In deep learning, style transfer involves inputting a "brown horse" image into a model to generate a "striped horse" image, without altering the horse's shape and size. This type of model is well-suited for handling noisy images input into a model to obtain their corresponding noise-free images, while preserving the shape and color of the river channel.

[0070] The automatic denoising method based on mismatched geological images of this invention comprises two generators with identical network structures and two discriminators with identical structures. The generators are built on the basis of convolutional autoencoders (DAEs), and the discriminators are structured similarly to an encoder. Since the main objective of this invention is to input a noisy image into the generator to obtain its corresponding noise-free image, noise is added to the input noisy image to prevent Gaussian noise-like interference from affecting the quality of the generated image, thereby enabling the generator to learn a more useful feature distribution.

[0071] We first input a noisy image X. noise Noise is added to it to obtain geological images with additional noise. The image is then input into the generator G, which is used to generate a noise-free image. A2B This results in a generated image that resembles a noise-free image. Then, we X... good Input to a generator G for generating noisy images. B2A This results in a generated image that resembles a noisy image. To enhance the generation effect of the two generators, we will generate images. Then input into generator G B2A This generates a looping image that resembles a noisy image. Then generate the image Then input into generator G A2B This results in a loop image that resembles a noise-free image. This process can be represented by the following formula:

[0072]

[0073]

[0074]

[0075]

[0076] Here, `add_noise()` represents adding additional noise to the original noisy input image. θg1 ,θ g2 These are the generator G A2B and generator G B2A Intermediate process parameters.

[0077] Then we will generate the image Compared with the original geological noise image X noise Input to discriminator D A We will compare the generated image with the original image. With noise-free image X good Input to discriminator D B Perform the original / generated distinction. This process can be represented by the following formula:

[0078]

[0079]

[0080] in, Discriminator D A The feature vectors output from the last layer correspond to X respectively. noise , Discriminator D B The feature vectors output from the last layer correspond to X respectively. good ,

[0081] The generator training for an automatic denoising method based on mismatched geological images mainly consists of three parts. First, we take the noise-free image X... good Input to generator G A2B Thus, the generated image is obtained. We will X-ray the noisy geological images noise Input to generator G B2A Obtain the generated image We minimize With image X good Distance between and minimization With image X noise The distance between them. This process can be represented as:

[0082]

[0083]

[0084] Where, loss ide_B With loss ide_A Representing the generator G respectively A2B With generator G B2A Image contrast loss.

[0085] The model uses a max-min game approach to train the generator and discriminator. The discriminator is fixed during generator training. For the generator G... A2B It expects to generate images Discriminator D B The result is true; for generator G B2A It expects to generate images Discriminator D A The result is determined to be true. This process can be represented as:

[0086]

[0087]

[0088] Where, loss gan_A2B With loss gan_B2A These are the generator G A2B With generator G B2A The loss of competitive training.

[0089] Furthermore, we minimize the loop image when training the generator. The distance between it and its corresponding original image. This process can be represented as:

[0090]

[0091]

[0092] Where, loss cyc_B With loss cyc_A It is the minimum contrast loss between the looped image from the two generators and the original image.

[0093] In summary, we can obtain the total loss of the generator. G It can be expressed by the following formula:

[0094] Loss G =α×(loss) ide_B +loss ide_A )+β×(loss gan_A2B +loss gan_B2A )+ρ×(loss cyc_B +loss cyc_A (15)

[0095] Among them, α=5, β=1, and ρ=10. Loss G This is the sum of the six losses mentioned above.

[0096] When training the discriminator, we fix the generator and first train the discriminator D. AIts purpose is to accurately identify the real image X. noise With generated image Then the discriminator D was trained. B Its purpose is to accurately identify the real image X. good With generated image Their training loss A With loss B It can be expressed by the following formula:

[0097]

[0098]

[0099] Figure 3 This shows a small geological image with noise and its corresponding denoised image. For example... Figure 3 As shown, by comparing small geological images with noise, we can obtain good, noise-free geological images. Good connectivity and integrity of the red river channels have been observed within each small image. The river channel morphology in the small images has been fully restored. Furthermore, we can observe that most of the horizontal and vertical lines and other cluttered noise in the blue non-river areas have also been removed, resulting in clean, non-river blue areas.

[0100] Figure 4 This shows the original noisy geological image and its corresponding denoising result. For example... Figure 4 As shown, it has been observed that despite the different river channel morphologies, the denoised results for each river morphology exhibit good connectivity and integrity. The surrounding noise is also largely removed. Multiple denoising results demonstrate that the model can achieve relatively accurate denoising results for different river morphologies.

[0101] Figure 5 Comparison images of the results of generating a model once and generating a model twice for the same geological image. Although Figure 3 Figure 4 The denoising results in the first generation already showed good river morphology, but we can observe that the image is still affected by other noise, causing some blurring. Therefore, we input the generated image back into the model to obtain a second generation result. We can observe that the second denoising result is significantly clearer than the first generation result. It is easy to see that the second denoising result fills in some small blue areas within the river and removes some small red spots outside the river area, thus achieving a better result.

[0102] Figure 6This is a comparison of the similarity between the results obtained by the unsupervised denoising method, the results of this patent, and the original noise-free image. For example... Figure 6 As shown, our denoising results outperform some unsupervised generative models in terms of cosine similarity and correlation coefficient, demonstrating the effectiveness of our model. It's worth noting that although our results are numerically superior to unsupervised results, the difference in similarity is not significant. However, compared to unsupervised methods, our method lacks the noise detection process and eliminates the need for multiple steps such as iterative filling. While the results are not significantly different, the operation is simpler and more convenient.

[0103] This invention proposes an automatic denoising method based on mismatched geological images. Current geological sedimentary models contain a large amount of messy noise, including horizontal and vertical stripes left over from model printing, as well as noise and regions that do not conform to geological laws due to errors in geological data. These noise points are randomly distributed and difficult to remove manually. Existing AI-based geological image denoising methods have high requirements for dataset preparation and the denoising steps are complex and difficult to perform. To address this problem, we propose an automatic denoising method based on mismatched geological images. First, we use multiple noisy geological images obtained from geological models using a cropping algorithm, along with a few noise-free images, and repeatedly cut them at different intervals to construct a mismatched noise-noise dataset with inconsistent river morphologies. This dataset does not require a one-to-one correspondence between noisy and noise-free images, thus simplifying the dataset construction process. Subsequently, based on the idea of ​​style transfer, we treat noisy images as the first class of images and noise-free images as the second class of images, thereby constructing an automatic denoising model based on a generative adversarial network. This model aims to output the corresponding denoising result for the input noisy geological image. This model simplifies the process of existing denoising methods by eliminating the need for prior noise detection. Denoising results for different river morphologies demonstrate that the model effectively removes incorrect noise around the river, thus ensuring the integrity and connectivity of the river channel.

[0104] Finally, the details of the above examples of the present invention are merely illustrative of the invention. Any modifications, improvements, and substitutions to the above embodiments by those skilled in the art should be included within the scope of protection of the claims of the present invention.

Claims

1. An automatic denoising method based on mismatched geological images, characterized in that, The method, used to remove incorrect noise around a river to ensure the integrity and connectivity of the river channel, includes the following steps: S1. Implement an automatic cropping algorithm for noisy geological images based on existing geological models, including: First, using geological software, the 3D geological model is converted into [0,1] codes in "csv" format; where 0 represents blue non-river areas and 1 represents red river channels. The "csv" code is then divided into multiple "csv" files of size 774×546. Given a 774×546 all-black image, color correction is performed on the all-black image based on the [0,1] codes in one of the obtained csv files. The codes in the csv file are iterated through according to rows 774 and columns 546. If the code is "1", the corresponding position in the all-black image is changed to white (255, 255, 255); if the code is "0", the pixel value at the corresponding position in the all-black image is not modified. This process yields a binarized noisy image. Then, the binarized image is color-filled to obtain a noisy geological image with red and blue contrasts. This process can be represented by the following formula: （1）； （2）； Among them, Z csv The "csv" file represents the model transformation, where z1, z2, and z3 represent the transformations from Z. csv The "csv" files corresponding to each segmented image; Represents the process of splitting a file; img black The first image represents a completely black image; img1, img2, and img3 represent the segmented geological images with noise; F represents the pixel modification process. Represents the process of color change; S2. Based on S1, multiple noisy geological images and noise-free images are cut at intervals to obtain a non-matching noisy-noise image dataset. S3. Construct an automatic denoising model for mismatched geological images; The automatic denoising model consists of two generators with the same network structure and two discriminators with the same structure; the generators are built on the basis of convolutional autoencoders (DAEs), and the discriminators are structured like an encoder. S4. Train the denoising model constructed in S3 based on the dataset in S2; The specific process of S4 is as follows: Generator training mainly consists of three parts; first, the noise-free image X is trained... good Input to generator G A2B Thus, the generated image is obtained. X-ray the noisy geological image. noise Input to generator G B2A Obtain the generated image ;minimize With image X good Distance between and minimization With image X noise The distance between them; this process can be represented as: （3）； （4）； in, and Representing the generator G respectively A2B With generator G B2A Image contrast loss; The model uses a max-min game approach to train the generator and discriminator; during generator training, the discriminator is fixed; and for the generator G... A2B The image to be generated Discriminator D B If the result is true, determine the generator G. A2B Loss of combat training For generator G B2A The image to be generated Discriminator D A If the result is true, determine the generator G. B2A Loss of combat training ; Minimize the cyclic image when training the generator ( , The distance between the image and its corresponding original image; this process can be represented as: （5）； （6）； in, and It is the minimum contrast loss between the looped image from the two generators and the original image; The total loss of the generator can be obtained. It can be expressed by the following formula: （7）； in, =5, =1, =10, This is the sum of the six losses mentioned above; When training the discriminator, with the generator fixed, the discriminator D is trained first. A Accurately identify the real image X noise With generated image Subsequently, the discriminator D was trained. B Accurately identify the real image X good With generated image .

2. The automatic denoising method based on mismatched geological images according to claim 1, characterized in that, The specific process of S2 is as follows: To ensure that there is no one-to-one correspondence between noisy and noiseless images in the constructed dataset, multiple noisy geological images of different shapes and a noiseless image of a different shape were simultaneously cut. The noisy geological images were cut into 128×128 pixels at intervals of 16, 32, and 64, resulting in 780 noisy images. The noiseless images were then cut at intervals and flipped to obtain 780 noiseless images. This non-matching noisy-noiseless image dataset thus eliminates the one-to-one correspondence between paired datasets.

3. The automatic denoising method based on mismatched geological images according to claim 1, characterized in that, The specific process of S3 is as follows: First, input a noisy image. Noise is added to it to obtain geological images with additional noise. The image is then input into a generator G for producing a noise-free image. A2B This results in a generated image that resembles a noise-free image. Subsequently, the original noise-free image X good Input to a generator G for generating noisy images. B2A This results in a generated image that resembles a noisy image. To enhance the generation effect of the two generators, the generated image will be... Then input into generator G B2A This generates a looping image that resembles a noisy image. ; The generated image is then input into generator G. A2B This results in a loop image that resembles a noise-free image. This process can be represented by the following formula: （8）； （9）； （10）； （11）； in, This means adding additional noise to the original noisy input image; , These are the generator G A2B and generator G B2A Intermediate process parameters; Then, the generated image Compared with the original geological noise image X noise Input to discriminator D A Distinguish between the generated image and the original image; convert the generated image... With noise-free image X good Input to discriminator D B Perform the original / generated distinction; this process can be represented by the following formula: （12）； （13）； in, , Discriminator D A The feature vectors output from the last layer correspond to X respectively. noise , ; , Discriminator D B The feature vectors output from the last layer correspond to X respectively. good , .

4. The automatic denoising method based on mismatched geological images according to claim 1, characterized in that, The method is implemented using PyTorch, and a new encoder, decoder and discriminator are redesigned based on a convolutional neural network. During training, a picture with a size of (batch, 3, 64, 64) is loaded each time, where batch is set to 5; the output image has a size of (batch, 3, 64, 64), and the output vector dimension of the discriminator is (batch, 1); in addition, the learning rate is set to 2x10 -4 , and the Adam optimizer is used for gradient update.

Images