Unsupervised automatic denoising method based on geological image visual features

Abstract

The application discloses an unsupervised automatic denoising method based on geological image visual features. A large number of noise points exist in geological model images, the noise points are small and dense, manual denoising is very complex and tedious, and therefore, noiseless image is not easy to obtain. In view of the problem, the application constructs a noiseless image data set which is expanded by using picture cutting and flipping only with a few noiseless images. From the perspective of visual features in the computer field, an unsupervised automatic noise point detection model is built. The model uses a newly designed convolutional autoencoder and a generative adversarial network to simultaneously control the generation of noiseless images from the image space and the feature space, and effectively restricts the intermediate process of encoding and decoding. The detection result shows that the model can fully detect incorrect noise points in the river channel, and 'horizontal bars' and'vertical bars' generated in the printing process. The denoising result obtained by the cyclic filling method is more in line with the geological law.

Description

Technical Field

[0001] This invention pertains to methods for geological noise reduction and industrial defect detection, 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. Researchers typically utilize computer graphics and visualization techniques to visually represent abstract geological data, creating realistic geological models. This allows for more efficient description of various geological information and effectively expresses the topological relationships between different geological phenomena, thereby rapidly improving the understanding of geological phenomena among professionals and enabling them to make accurate predictions. 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, thus enhancing the selectivity of geological models. Filtersim uses filters for pattern classification and utilizes 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.

[0003] 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.

[0004] In the field of computer science, the denoising process of geological images can be transformed into the process of finding outliers. Anomaly detection has been extensively studied, and researchers generally use traditional techniques such as clustering, nearest neighbor methods, and single-class classification. For example, single-class SVM can be used to find the discriminative boundaries around positive samples. Clustering methods such as K-Means and Gaussian Mixture Models (GMM) have also been applied to build detailed profiles of normal data to identify anomalies. However, these methods often perform poorly when dealing with high-dimensional data. To address the detection problem of complex data, reconstruction-based methods have been proposed, whereby anomalies cannot be accurately represented and reconstructed using models learned only on normal data. For example, Principal Component Analysis (PCA) has been used to learn the feature distribution of normal samples; sparse representation methods jointly learn a data dictionary and the sparse representation features of normal samples to better reconstruct positive samples. However, the limited feature representation restricts the performance of these methods.

[0005] In recent years, several anomaly detection algorithms based on convolutional autoencoder (AE) networks have emerged for deep learning tasks in computer vision. Deep AE-based convolutional autoencoders are used to model training samples; researchers have proposed jointly encoding vectors and reconstruction errors in AEs. Although reconstruction-based methods have achieved significant results, their performance is limited by insufficient latent space design. To address this issue, Generative Adversarial Networks (GANs) have been applied to anomaly detection. GANs can simulate complex high-dimensional distributions, especially images, through a mini-maximum game process, and the models can successfully generate data with excellent performance. Schlegl et al. proposed AnoGAN, which uses a similar convolutional structure to DCGAN to build a generator trained only on normal images, and then calculates anomaly scores based on the reconstructed images and discriminator feature maps to discover anomaly markers in medical images. Later, to reduce the complexity of the mapping from images to the 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. Summary of the Invention

[0006] The purpose of this invention is to replace manual denoising, solve the problem of uneven distribution between noisy and non-noisy images, and build a computer-automated denoising model to solve the problem of automatic denoising of geological images.

[0007] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0008] S1. Construct a noise-free geological image dataset by slicing to improve the problem of very few noise-free images.

[0009] S2. An unsupervised geological image noise detection model is built based on generative adversarial networks.

[0010] S3. Train an unsupervised geological image noise detection model based on the noise-free image dataset of S1.

[0011] S4. Based on the model trained in S3, noise detection is performed on the geological image. Then, based on the noise detection results, pixel-level denoising and cyclic denoising methods are used to denoise the geological image.

[0012] The noise-free image dataset of this invention is constructed based on a segmentation method and expanded using flipping. Since geological images are mostly noisy, a limited number of noise-free images are difficult to use for model training. The generator in the model only needs to learn the connectivity and integrity of the red region (river channel), without special attention to features such as river channel shape. Therefore, the segmentation method does not affect the experimental results. Because the initial noise-free images have "black borders," to avoid them interfering with the training of the detection model, we first use an image preprocessing module to remove the black borders around the original images.

[0013] Because the black pigment value is small, close to 0, we set a threshold and determined the position without black borders by comparing pixels one by one. If the value is greater than the threshold, it means that the position is the initial position without black borders; if the value suddenly falls below the threshold, it means that it is the end position without black borders. We recorded the positions of the positions without black borders horizontally and vertically. The original noise-free image was cropped based on the four positions to obtain a noise-free image without black borders. This process can be represented by the following formula.

[0014]

[0015]

[0016] Here, flag is an intermediate value set during the calculation process, initialized to False, threshold is the threshold value set, the threshold size is 25, hMax[i] and wMax[i] are the maximum pigment values ​​of each row and each column in the original geological image, and position_h and position_w are the initial and end position pairs of the image without black borders that we need to record.

[0017] We segment the preprocessed noise-free images at intervals, dividing them into 128×128 noise-free images at intervals of 16, 32, 64, 128, and 256. Then, we flip the images to obtain the noise-free image dataset.

[0018] This invention trains with a positive sample dataset (noise-free images) without the need for negative samples (noisy images), which solves both the problem of imbalanced distribution between the two and the problem of a small number of noise-free images.

[0019] The unsupervised geological image noise detection model in this invention is based on a generative adversarial network (GAN). It consists of a generator and a discriminator; adversarial training between the generator and discriminator helps the generator continuously learn the feature distribution of positive samples. The generator is primarily based on a convolutional autoencoder (AE), but here we use an improved convolutional denoising autoencoder (DAE). More useful feature distributions are learned by reconstructing input data containing additional noise.

[0020] Given X as the domain of a data sample, and Z as the domain of the encoding, let f e1 (·): X→Z represents the encoder, f d (·): Z→X represents the decoder. It adds noise to the noise-free image X. The resulting image is then labeled with noise. The input is fed into the encoder, and after passing through multiple convolutional layers, a noise-free encoded vector is obtained. vector The input is fed into the decoder to obtain the reconstructed noise-free image. Discriminator f D It is an image encoding process that reconstructs a noise-free image. The corresponding noise-free image X is input into the discriminator f. D In this process, their discriminant vectors are obtained. Z D The discriminator determines whether an image is real or fake based on the obtained vectors. The main function of the discriminator is to assist the generator training to obtain excellent reconstructed images. This process can be represented by the following formula:

[0021]

[0022]

[0023]

[0024] Where, θ e1 With θ d These are encoder f e1 and decoder f d The parameters. add noise(.) It involves adding other noise functions. θ D1 With θ D2 These are the discriminators f D In judging the reconstructed noise-free image The parameters for the noise-free image X.

[0025] This invention incorporates encoder f in the generator. e1 Encoders f with the same network structure e2 (·). The original generator only applies loss constraints in the image space, often neglecting network training in the latent space. This leads to the encoder's extraction process and the decoder's decoding process being beyond our control. Therefore, another encoder is added to encode the reconstructed noise-free image, allowing for a comparison between the encoded features of the noise-free image and the encoded features of the reconstructed image, thus ensuring constraints in the feature space. This process can be represented as:

[0026]

[0027] Where, θ e2 Indicates encoder f e2 Parameters during the encoding process. Represents the reconstructed image The coding features.

[0028] Furthermore, due to limitations in the network structure, the quality of the generated images is insufficient. Therefore, during testing, the generated images obtained from our input noisy image were... The difference between the generated image and the input image is insufficient to directly represent the denoising result. Therefore, we subtract the input image from the generated image to obtain the difference image X. difference This refers to the noise detection result. The noise detection process can be represented by the following formula:

[0029]

[0030] Among them, sub R ,sub G ,sub B This represents the subtraction operation between the input noisy image and its corresponding generated image on the R, G, and B channels, respectively. binarized Binarization operation representing the result.

[0031] The main training process of the unsupervised geological image noise detection model in this invention is as follows:

[0032] First, the loss function used in training mainly includes the generator loss and the discriminator loss, and adversarial training is performed using a minimax game approach. For the basic generative adversarial network (GAN), we retain the loss function of the GAN network: when training the generator, the discriminator is fixed, allowing the generator-generated images to fool the discriminator as much as possible; when training the discriminator, the generator is fixed, allowing the discriminator to distinguish as much as possible whether an image is generated by the generator or the original image. This process can be expressed by the following formula:

[0033]

[0034]

[0035] Where, loss adv The adversarial loss representing the generator aims to make The encoding that is identified as the original image. D This represents the total loss of the discriminator.

[0036] Besides the generator's adversarial loss adv In addition, the generator also includes contrast loss between images and contrast loss between features. The image loss is the loss between the noise-free image X and the reconstructed image obtained after feeding it into the generator. The reconstructed noisy image is compared with its corresponding noise-free image, and the loss between the two is minimized during training. The loss between features is minimized by the encoder f. e1 With encoder f e2 The extracted features. This process can be represented as:

[0037]

[0038]

[0039] loss G =α*loss adv +β*loss img +ρ*loss z (12)

[0040] Where, loss img The loss represents the contrast loss between images. z The contrast loss represents the loss between features in the latent space. The overall loss of the generator is... G It includes three losses. α and ρ are initially set to 1 and β is initially set to 50.

[0041] 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, 128, 128) at a time, where the batch size is set to 5, and the encoder f... e1 and f e2 The output encoded vector size is (batch, 512, 8, 8), and the final output of the discriminator is (batch, 1, 14, 14). Additionally, we set the epoch size to 100 and the learning rate to 1×10⁻⁶. -4 The Adam optimizer is used for gradient updates.

[0042] To alleviate the pressure of model testing, this invention employs a pixel-level denoising method to initially remove some noise. The original geological image contains fine and complex noise. Among the three types of noise, the noise within the red river channel needs to be detected by the model, but the small red spots appearing outside the red area and in the large blue area can be removed without a model. This method is based on contour detection from the field of natural image processing, delineating the contours of non-river areas (large blue areas) to fill and remove noise. The method consists of six steps. First, the original geological image is binarized, converting it to black and white for easier subsequent processing. Second, the binarized image is denoised using Gaussian filtering to smooth the image. Then, the image undergoes flooding and opening / closing operations to initially determine the contours. Finally, the desired region contours are delineated and filled with color.

[0043] This invention employs a cyclic filling method for automatic noise removal. In the original noisy image, noise includes small blue spots within red areas, including horizontal and vertical stripes; these noise spots need to be automatically restored to red by the computer. Additionally, small red horizontal stripes within blue areas are also noise that needs to be removed, automatically changing the image to pure blue. Therefore, based on the noise detection results, this method modifies the pixel values ​​of the noise at the corresponding noise locations in the original noisy image to either blue or red. By repeating this process, automatic noise removal can be achieved.

[0044] Compared with existing technologies, the beneficial effects of this invention are:

[0045] 1. This invention proposes a deep learning model based on generative adversarial networks and autoencoder networks, which automatically detects noise in geological sedimentary model images according to the unsupervised approach.

[0046] 2. Using only a small number of noise-free images, a noise-free image dataset was created using image cropping and image flipping techniques, avoiding the hassle of manual labeling; and avoiding the intervention and labeling of noisy images. Attached Figure Description

[0047] Figure 1 This is the original geological noise image.

[0048] Figure 2 This is a schematic diagram of the structure of an unsupervised geological image noise detection model.

[0049] Figure 3 A schematic diagram illustrating the process of preparing a noise-free image dataset.

[0050] Figure 4 The images show the original geological image with noise and the image after denoising.

[0051] Figure 5This is a noise detection image obtained by an unsupervised automatic denoising method based on the visual features of geological images.

[0052] Figure 6 The image is the result of five noise detection operations performed using an unsupervised automatic denoising method based on the visual features of geological images. Detailed Implementation

[0053] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0054] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0055] Figure 1 Images containing geological noise generated during geological exploration. For example... Figure 1 As shown, due to certain errors in the predicted geological internal data, the geological images obtained from the modeling produce some noise that does not conform to geological laws (small red spots inside the large blue area); in addition, some essential operations during the printing of the geological model will cause the resulting geological images to produce incorrect noise such as small "horizontal bars" and "vertical bars".

[0056] Figure 2 This is a schematic diagram of the structure of an unsupervised geological image noise detection model. (Example:) Figure 2 As shown, the base network of the unsupervised geological image noise detection model is a generative adversarial network (GAN). It consists of a generator and a discriminator; adversarial training between the generator and discriminator helps the generator continuously learn the feature distribution of positive samples. The generator is primarily based on a convolutional autoencoder (AF), but here we use an improved convolutional denoising autoencoder (DAF). More useful feature distributions are learned by reconstructing input data containing additional noise.

[0057] Given X as the domain of a data sample, and Z as the domain of the encoding, let f e1 (·): X→Z represents the encoder, f d (·): Z→X represents the decoder. It adds noise to the noise-free image X. The resulting image is then labeled with noise. The input is fed into the encoder, and after passing through multiple convolutional layers, a noise-free encoded vector is obtained. vector The input is fed into the decoder to obtain the reconstructed noise-free image. Discriminator f D It is an image encoding process that reconstructs a noise-free image. The corresponding noise-free image X is input into the discriminator f. D In this process, their discriminant vectors are obtained. Z DThe discriminator determines whether an image is real or fake based on the obtained vectors. The main function of the discriminator is to assist the generator training to obtain excellent reconstructed images. This process can be represented by the following formula:

[0058]

[0059]

[0060]

[0061] Where, θ e1 With θ d These are encoder f e1 and decoder f d The parameters. add noise(.) It involves adding other noise functions. θ D1 With θ D2 These are the discriminators f D In judging the reconstructed noise-free image The parameters for the noise-free image X.

[0062] This invention incorporates encoder f in the generator. e1 Encoders f with the same network structure e2 (·). The original generator only applies loss constraints in the image space, often neglecting network training in the latent space. This leads to the encoder's extraction process and the decoder's decoding process being beyond our control. Therefore, another encoder is added to encode the reconstructed noise-free image, allowing for a comparison between the encoded features of the noise-free image and the encoded features of the reconstructed image, thus ensuring constraints in the feature space. This process can be represented as:

[0063]

[0064] Where, θ e2 Indicates encoder f e2 Parameters during the encoding process. Represents the reconstructed image The coding features.

[0065] Furthermore, due to limitations in the network structure, the quality of the generated images is insufficient. Therefore, during testing, the generated images obtained from our input noisy image were... The difference between the generated image and the input image is insufficient to directly represent the denoising result. Therefore, we subtract the input image from the generated image to obtain the difference image X. difference This refers to the noise detection result. The noise detection process can be represented by the following formula:

[0066] X difference =f binarized(max(sub R ,sub G ,sub B (5)

[0067] Among them, sub R ,sub G ,sub B This represents the subtraction operation between the input noisy image and its corresponding generated image on the R, G, and B channels, respectively. binarized Binarization operation representing the result.

[0068] The main training process of the unsupervised geological image noise detection model is as follows:

[0069] First, the loss function used in training mainly includes the generator loss and the discriminator loss, and adversarial training is performed using a minimax game approach. For the basic generative adversarial network (GAN), we retain the loss function of the GAN network: when training the generator, the discriminator is fixed, allowing the generator-generated images to fool the discriminator as much as possible; when training the discriminator, the generator is fixed, allowing the discriminator to distinguish as much as possible whether an image is generated by the generator or the original image. This process can be expressed by the following formula:

[0070]

[0071]

[0072] Where, loss adv The adversarial loss representing the generator aims to make The encoding that is identified as the original image. D This represents the total loss of the discriminator.

[0073] Besides the generator's adversarial loss adv In addition, the generator also includes contrast loss between images and contrast loss between features. The image loss is the loss between the noise-free image X and the reconstructed image obtained after feeding it into the generator. The reconstructed noisy image is compared with its corresponding noise-free image, and the loss between the two is minimized during training. The loss between features is minimized by the encoder f. e1 With encoder f e2 The extracted features. This process can be represented as:

[0074]

[0075]

[0076] loss G =α*loss adv +β*lossimg +ρ*loss z (10)

[0077] Where, loss img The loss represents the contrast loss between images. z The contrast loss represents the loss between features in the latent space. The overall loss of the generator is... G It includes three losses. α and ρ are initially set to 1 and β is initially set to 50.

[0078] Figure 3 The image processing steps for creating a noise-free image dataset. For example... Figure 3 As shown, the noise-free image dataset is constructed based on a cutting method and augmented using flipping. Since geological images are mostly noisy, the limited number of noise-free images is insufficient for model training. The generator in the model only needs to learn the connectivity and integrity of the red region (river channel), without special attention to features such as river channel shape. Therefore, the cutting method does not affect the experimental results. Because the initial noise-free images have "black borders," to avoid them interfering with the training of the detection model, we first remove the black borders around the original images using an image preprocessing module.

[0079] Because the black pigment value is small, close to 0, we set a threshold and determined the position without black borders by comparing pixels one by one. If the value is greater than the threshold, it means that the position is the initial position without black borders; if the value suddenly falls below the threshold, it means that it is the end position without black borders. We recorded the positions of the positions without black borders horizontally and vertically. The original noise-free image was cropped based on the four positions to obtain a noise-free image without black borders. This process can be represented by the following formula.

[0080]

[0081]

[0082] Here, flag is an intermediate value set during the calculation process, initialized to False, threshold is the threshold value set, the threshold size is 25, hMax[i] and wMax[i] are the maximum pigment values ​​of each row and each column in the original geological image, and position_h and position_w are the initial and end position pairs of the image without black borders that we need to record.

[0083] We segment the preprocessed noise-free images at intervals, dividing them into 128×128 noise-free images at intervals of 16, 32, 64, 128, and 256. Then, we flip the images to obtain the noise-free image dataset.

[0084] This invention trains with a positive sample dataset (noise-free images) without the need for negative samples (noisy images), which solves both the problem of imbalanced distribution between the two and the problem of a small number of noise-free images.

[0085] Figure 4 For comparison of the denoised results with the original geological image. Figure 4 The first and third rows are the original geological images with noise, while the second and fourth rows are the denoising results after our model detected the noise. For example... Figure 4 As shown, the proposed method can remove noise within the red channel, and red spots within the large blue area are also removed. This fully demonstrates the effectiveness of the invention.

[0086] Figure 5 This is a comparison between the noise detection results and the original geological image. Figure 5 The first and third rows are images with noise, while the second and fourth rows are noise-detected images obtained from the first detection by the trained model. As can be seen from the images, the model detected most of the incorrect noise present within the river channel, and also detected the "horizontal" and "vertical" lines generated during the printing process. This fully demonstrates the model's excellent noise detection performance.

[0087] Figure 6 These are the results of five noise detection iterations obtained using an unsupervised automatic denoising method based on visual features of geological images. Figure 6 As can be seen from the five detection results, the noise detected by the model decreases with each detection until the fifth detection, when the noise is similar to the outline of the river in the image. Therefore, it can be considered that the noise has been basically detected.

[0088] This invention proposes an unsupervised automatic denoising method based on visual features of geological images. Geological images contain complex and dense noise, making manual detection and removal time-consuming and labor-intensive. To address this problem, the model adopts an unsupervised approach, creating a noise-free image dataset. A newly designed convolutional autoencoder and generative adversarial network are used to control the generator's reconstruction of the noise-free images from both the image space and feature space perspectives. Detection results show that the model can detect most of the noise in the original noisy images, and the restored geological images are free of incorrect noise, better reflecting the principles of geological exploration. In the future, we will continue to research more powerful detection methods, combining computer science with geology to apply geological knowledge to the computer field.

[0089] 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 unsupervised automatic denoising method based on geological image visual features, characterized in that, The method comprises the following steps: S1. Constructing a non-noise geological image dataset by slicing to improve the problem of few non-noise images; S2. Building an unsupervised geological image noise detection model based on a generative adversarial network; S3. Training the unsupervised geological image noise detection model based on the non-noise image dataset of S1; training the unsupervised geological image noise detection model according to the total loss of the generator; wherein the total loss of the generator is represented as: loss G = a x loss adv + b x loss img + p x loss z (12) ； ； ； ； ； ； ； where θ e1 and θ d are parameters of the encoder f e1 and the decoder f d respectively; add noise (·) is an additive noise function; θ D1 and θ D2 are parameters of the discriminator f D when discriminating the reconstructed noise-free image and the noise-free image X; θ e2 represents the parameters of the encoder f e2 during the encoding process; represents the encoding feature of the reconstructed image ; is the noise-free encoding vector; is the reconstructed noise-free image input to the discriminator f D to obtain the discriminative vector; Z D is the discriminative vector obtained by inputting the reconstructed noise-free image corresponding noise-free image X to the discriminator f D ; f e1 (·) is the encoder; f e2 (·) is an encoder having the same network structure as the encoder f e1 (·); loss adv represents the adversarial loss of the generator, which aims to make the encoding be discriminated as the original image; loss img represents the contrast loss between images, loss z represents the contrast loss between features in the latent space, and the overall loss loss G of the generator contains three losses, and α and ρ are initially set to 1 and β is initially set to 50; The image loss is the reconstructed image obtained after the noise-free image X is fed into the generator. The reconstructed noisy image is compared with its corresponding noise-free image, and the loss between the two is minimized during training; the loss between features is minimized by the encoder f. e1 With encoder f e2 Extracted features; S4. Noise detection of the geological image based on the trained model of S3; and removing noise from the geological image using a pixel-level denoising method and a cyclic denoising method according to the noise detection result.

2. The unsupervised automatic denoising method based on geological image visual features according to claim 1, characterized in that, The specific process of S1 is as follows: The non-noise image dataset is constructed based on the cutting method and expanded using flipping; including: First, remove the black border around the original image through the image preprocessing module; Because the black color value is small and close to 0, a threshold is set to determine the position of the non-black border by comparing each pixel; According to the four positions, the original non-noise image is cropped to obtain a non-noise image without a black border; this process can be represented by the following formula: ； ； Wherein, flag is an intermediate value set during calculation, initialized as False, threshold is the set threshold, the threshold size is 25, hMax[i] and wMax[i] are the maximum color value of each row and each column of the original geological image, and position_h and position_w are the initial and ending position pairs of the image without black border that need to be recorded. The preprocessed non-noise image is cut at intervals, cut into 128x128 non-noise images at intervals of 16, 32, 64, 128, 256, etc., and the image is flipped to obtain a non-noise image dataset. 3.The method of claim 1, wherein, The specific process of S2 is as follows: The basic network of the unsupervised geological image noise detection model is a generative adversarial network; the generative adversarial network comprises a generator and a discriminator; the generator and the discriminator are trained in opposition to promote the generator to continuously learn the feature distribution of the positive sample; the generator is mainly based on a convolutional autoencoder (AE), and an improved convolutional denoising autoencoder (DAE) is used as the generator here; the generator learns a more useful feature distribution by reconstructing the input data containing other noise; Given X is the domain of data samples, Z represents the domain of encoding; let f e1 (·):X→Z represents the encoder, f d (·):Z→X represents the decoder; add other noise to the noise-free image X; Images with noise The input is fed into the encoder, and after passing through multiple convolutional layers, a noise-free encoded vector is obtained. , will vector The input is fed into the decoder to obtain the reconstructed noise-free image. Discriminator f D It is an image encoding process that reconstructs a noise-free image. The corresponding noise-free image X is input into the discriminator f. D In this process, the discriminant vector is obtained. Z D The discriminator determines whether an image is real or fake based on the obtained vectors. The main function of the discriminator is to assist the generator in training to obtain excellent reconstructed images. Addition with generator in encoder f e1 Encoder f with same network structure e2 (·) At the time of testing, for the input image with noise, the generated image Cannot be directly subtracted as a de-noising result, the generated image and the input image, to get the difference image X difference , that is, the detection result of the noise; the noise detection process can be expressed by the following formula: X difference = f binarized (max(sub R ,sub G ,sub B )) (7) wherein, sub R , sub G , sub B denote the subtraction operation of the input noisy image and its corresponding generated image in R, G, B channels respectively, f binarized represents the binarization operation of the result. 4.The method of claim 1, wherein, The specific process of S3 is as follows: The main training process of the unsupervised geological image noise detection model in the application is as follows: First, the loss function mainly includes the generator loss and the discriminator loss, and the maximum minimum game method is used for adversarial training; for the basic generative adversarial network, the loss function of the GAN network is retained, that is, when training the generator, the discriminator is fixed, so that the image generated by the generator can deceive the discriminator; When training the discriminator, the generator is fixed, so that the discriminator can determine whether the image is the image generated by the generator or the original image; ； where loss D represents the total loss of the discriminator; represents the discriminative vector; Z D is the reconstructed noise-free image The corresponding noise-free image X is input to the discriminator f D to obtain the discriminative vector. 5.The method of claim 1, wherein, The specific process of S4 is as follows: Part of the noise is preliminarily removed by using a pixel-level denoising method; The pixel-level denoising method fills and removes the noise by circumscribing the contour of the non-river area representing the large blue area, and the pixel-level denoising method comprises the following steps: firstly, the original geological image is binarized to make the color black and white, which is convenient for subsequent processing; secondly, the binarized image is subjected to Gaussian filter denoising for smoothing the image; then, the image is subjected to a flood fill algorithm and an open-close operation to preliminarily determine the contour; finally, the required area contour is circumscribed, and the color is filled. The cycle filling method is used to automatically remove the noise points; wherein, according to the obtained noise point detection result, the pixel value of the noise point corresponding to the red horizontal bar of the blue area is modified to blue or the pixel value of the noise point corresponding to the blue spot inside the red area is modified to red at the corresponding noise point position of the original noise point image, and the process is repeated to automatically remove the noise points.

Images