Geological model image denoising method and system, storage medium and equipment

Abstract

The embodiment of the invention discloses a geologic model image denoising method and system, a storage medium and equipment, and the method comprises the steps: firstly obtaining a data set with noisy point-noisy point-free image pairing, avoiding the complex work of noisy point marking and the like, saving human resources, training a noisy point detection model obtained based on the data set, and improving the noise detection efficiency. It is guaranteed that the model plays a role in obtaining the noisy point position on an input noisy point image in the later period, the capacity of the model for capturing noisy point distribution is enhanced, the model is more suitable for detection of geological noisy points, preliminary removal of the noisy points is achieved through a pixel-level denoising method, then detection of the remaining noisy points is conducted through the noisy point detection model, and the method is provided for improving the noisy point removal accuracy. The pixel-level denoising model and the noisy point detection model are combined, cyclic multiple detection can be carried out, errors caused by missing detection of the model are reduced, and therefore denoising comprehensiveness is guaranteed, a clear and complete result image of a river channel is generated, and judgment of the underground form and prediction of an oil layer in the later period are facilitated.

Description

Technical Field

[0001] The invention belongs to the technical field of geological model image denoising, and relates to a geological model image denoising method, system, storage medium and equipment. Background Art

[0002] Geological modeling technology has gradually developed from two-point geostatistical modeling to multi-point geostatistical modeling. Two-point geostatistical modeling describes spatial characteristics through variograms. Deterministic models are established using the Kriging method and optional and equally likely geological models are established using random simulation methods. However, two-point geostatistical modeling cannot accurately characterize complex spatial structures and the geometric morphology of the target body. Many scholars have studied multi-point geostatistical models. The Snesim algorithm uses training images instead of the variogram between two points to reproduce the reservoir configuration of complex structures. The Simpat algorithm adds similarity calculations to compare the similarity between data events at the estimated point and the training images, thereby enhancing the selectivity of geological models. The Filtersim algorithm uses filters for pattern classification and uses overall patterns instead of single patterns to reduce computer burden. The Growthsim algorithm uses optimized spatial simulation and a unique growth-based algorithm to reproduce patterns based on Snesim. However, these modeling techniques are still insufficient to provide a reasonable three-dimensional geological model.

[0003] With the development of artificial intelligence technology, geostatistical modeling has been pushed to the stage of intelligent modeling. Current work focuses on using generative adversarial networks to consider more complex geological patterns and statistical information, thereby generating more realistic geological models. Specifically, the spatial generative adversarial neural network SGAN is used to reduce the dimension of probabilistic inversion through geostatistical inversion of training images, ensuring unconditional realization of 3D. The improved adversarial generative network WGAN is used to train the model, which directly generates realistic geological models through a small number of random parameters. However, the modeling method based on deep learning requires a large number of training samples. Therefore, some artificial neural network prediction methods based on nonlinear mapping have gradually replaced the data synthesis method based on expert experience. Whether it is traditional geological modeling methods or novel deep learning modeling, inaccurate data will lead to noise in the constructed geological model. They will destroy the integrity and connectivity of the river channel in the model, thereby interfering with our judgment of the underground morphology and affecting the prediction of oil layers. Summary of the invention

[0004] The purpose of the present invention is to solve the problem that a large number of irregular noise points are generated in a specific geological model due to incomplete geological data, data errors and model printing operations in the prior art. Since the noise points are complexly distributed, these noise points are difficult to remove manually, and the number of existing noise-free images is small, and the noise points are difficult to mark manually. A geological model image denoising method, system, storage medium and device are provided.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for denoising a geological model image comprises the following steps:

[0007] Construct a paired dataset of noisy images and noise-free images, and build a noise detection model based on the paired dataset;

[0008] Construct a pixel-level denoising model, remove some noise points in the original geological image based on the pixel-level denoising model, and obtain an image after removing some noise points;

[0009] The noise detection model is used to obtain the generated result of the image after removing some noise points. The generated result is compared with the image after removing some noise points to obtain the noise point position. The remaining noise points are removed based on the obtained noise point position to obtain the denoised and repaired image.

[0010] A further improvement of the present invention is:

[0011] The step of constructing a paired dataset of noisy images and noise-free images comprises the following steps:

[0012] Acquire geological images with noise and geological images without noise, and preprocess the geological images with noise and geological images without noise so that each pair of image channels correspond to each other, and obtain image pairs corresponding to the channel morphology;

[0013] By cutting with different intervals γ, we get the image pair dataset.

[0014] It also includes training the noise detection model:

[0015] The model minimizes the noise-free image X good Reconstruction of noisy images Make the generator according to X good Learn X noise The location of the medium noise point;

[0016] Next X good and images The contrast loss in the latent space is used to strengthen the generator's attention to the difference between the two from the feature level;

[0017] Encoder e2 with the same structure as encoder e1 is used to extract the noise-free image X good The encoding feature Z2 is compared with the vector Z1 obtained by encoder e1 to bring the encoding process of the noise point and the noise-free point distribution closer:

[0018]

[0019] Loss z (Z1,Z2)=||Z1-Z2||1 (2)

[0020] Among them, Loss img and Loss z They are the contrast losses at the image level and feature level respectively;

[0021] Generator Loss G for:

[0022]

[0023] Among them, α, β, ρ are weight parameters, β = 50, α = ρ = 1; loss GAN is the original GAN ​​generator adversarial loss, which represents the minimum The distance from 1 makes the discriminator unable to distinguish;

[0024] The discriminator is trained by feature-level contrast loss, and the features of the last layer of the discriminator are and Used for loss judgment, to bring the discriminant features closer distance from 0 and minimize the feature Distance from 1:

[0025]

[0026] Among them, Loss D represents the overall loss of the model discriminator, Represents the noisy image output by the discriminator The eigenvector of Represents the noise-free image X output by the discriminator good The feature vector of .

[0027] The method of removing some noise points existing in the original geological image based on the pixel-level denoising model includes:

[0028] Binarization of the denoised geological image is performed to obtain a black and white image;

[0029] The obtained black and white image is processed by a Gaussian filter to obtain a weighted average value of pixels in the neighborhood, and the weighted average value of pixels in the neighborhood is used to replace the central pixel value to obtain a Gaussian filtered image;

[0030] Based on the Gaussian filtered image, the Flood Fill algorithm is used to highlight the important areas in the river.

[0031] Based on the image obtained after Flood Fill, the non-river channel area in the geological image is determined and circled through opening and closing operations, the circled area is filled with color, and the noise points in the area are modified to the corresponding correct pixel values ​​to achieve preliminary denoising.

[0032] The Flood Fill algorithm is used to color-enhance important areas in the river, including:

[0033] The seed pixel value can be set to 150, and the newly assigned pixel value of the flood area can be set to (255,255,255). The difference between the pixel values ​​before and after is about 2-3. The color of the noise points and the discontinuities in the river channel in the flooded image is darker.

[0034] The opening and closing operations to determine the non-river channel area in the geological image include:

[0035] Use the opening operation to discard the independent small red blocks inside the blue area and only select the overall coherent areas;

[0036] The closing operation is used to fill the blue gaps inside the red area to ensure the connectivity of the large red area, and then determine the outline of the blue area, that is, to delineate the non-river area.

[0037] The process of comparing the image after removing some noise points with the corresponding model generation result includes:

[0038] Given a noisy image X noise , after pixel-level denoising, we get the pixel-level image X Pixel ;

[0039] X pixel Input into the noise detection model to detect the remaining noise points and obtain the noise detection image X detect

[0040] According to X detect Find the corresponding noise points on the noise image and fill them with red or blue colors to obtain the noise repair image X. result1 .

[0041] The noise detection image X detect The acquisition process is:

[0042] Subtract the generated image from the geological image input to the model on the three channels (R, G, B) to obtain the detection result X of the location with noise detect :

[0043]

[0044] Among them, Abs R , Abs G , Abs B represents the subtraction operation between the geological image and the generated result on (R, G, B), f threshold The function is to binarize the result of noise detection.

[0045] A geological model image denoising system comprises a noise detection model building module, a first noise removal module and a second noise removal module;

[0046] A noise detection model building module is used to build a paired dataset of noisy images and noise-free images, and to build a noise detection model based on the paired dataset;

[0047] The first noise removal module is used to construct a pixel-level denoising model, remove some noise points in the original geological image based on the pixel-level denoising model, and obtain an image after removing some noise points;

[0048] The second noise removal module obtains the generated result of the image after removing part of the noise through the noise detection model, compares the generated result with the image after removing part of the noise to obtain the noise position, removes the remaining noise based on the obtained noise position, and obtains the denoised and repaired image.

[0049] A terminal device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any method described in the present invention when executing the computer program.

[0050] A computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the steps of any method described in the present invention.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] The embodiment of the present invention discloses a geological model image denoising method. First, a data set of paired images with noise and without noise is obtained, so as to avoid complex work such as noise marking and save human resources. A noise detection model obtained based on the data set is trained to ensure that the later model can obtain the position of noise for the input noise image, enhance the ability of the model to capture the distribution of noise, and make the model more suitable for the detection of geological noise. The pixel-level denoising method is used to achieve preliminary removal of noise, and then the noise detection model is used to detect the remaining noise. In order to improve the accuracy of noise removal, the pixel-level denoising model is combined with the noise detection model, and multiple detections can be performed in a cycle to reduce errors caused by missed detections of the model, thereby ensuring the comprehensiveness of denoising to generate a clear and complete result image of the river channel, which is helpful for the later judgment of underground morphology and prediction of oil layers. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0054] Figure 1 A flowchart for constructing a noise-noise image pair data set of the present invention;

[0055] Figure 2 Schematic diagram of the noise detection model and cyclic denoising method of the present invention;

[0056] Figure 3 is a flow chart of the pixel level denoising method of the present invention;

[0057] Figure 4 This is the noise detection result disclosed in the present invention.

[0058] Figure 5 These are the denoising results of various rivers disclosed in the present invention and the noise point detection results obtained in each cycle.

[0059] Figure 6 This is the denoising result of different river channels obtained by the pixel-level denoising method of the present invention.

[0060] Figure 7 This is a comparison result diagram of the multi-scale contrastive learning network of the present invention with unsupervised AE and GAN as well as supervised AE and GAN under five standards.

[0061] Figure 8 The figure shows a comparison between the noise detection results and denoising results obtained by an unsupervised automatic denoising method based on visual image features and the results obtained by the present invention. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0063] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0065] In the description of the embodiments of the present invention, it should be noted that if the terms "upper", "lower", "horizontal", "inner", etc. indicate an orientation or positional relationship based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship in which the product of the invention is usually placed when in use, it is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0066] In addition, if the term "horizontal" appears, it does not mean that the component must be absolutely horizontal, but can be slightly tilted. For example, "horizontal" only means that its direction is more horizontal than "vertical", which does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0067] In the description of the embodiments of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", and "connect" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal connection of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0068] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0069] See also Figure 1 The embodiment of the present invention discloses a geological model image denoising method, which trains a multi-scale contrast learning network based on a noise-noise-free image pair data set, and jointly learns the characteristic distribution of the river channel in the image in the feature space and the image space. A pixel-level denoising method is constructed based on the idea of ​​contour detection to preliminarily remove noise points in some non-river areas; a cyclic denoising method is constructed based on the idea of ​​recursion to remove noise points in most geological images by multiple noise detection and multiple pixel value filling, thereby ensuring the integrity of denoising. This method can remove noise points in most geological images, thereby ensuring the integrity and connectivity of the river channel, and providing assistance for the prediction of oil and gas reservoirs.

[0070] The following steps are involved:

[0071] Step 1: Build a paired dataset of noisy images and noise-free images, and build a noise detection model based on the paired dataset;

[0072] In the embodiment of the present invention, a few paired geological images and their corresponding noise-free images are preprocessed so that each pair of images has a one-to-one correspondence with the river channel. Then, a noise-noise-free image pair data set is obtained by image cutting and flipping, and a noise detection model is constructed based on the data set. The model is trained and learned from the image space and feature space to learn the distribution of river channel features, specifically:

[0073] See also Figure 1 , construct a paired dataset of noisy images and noise-free images for comparative training:

[0074] Among them, the geological images with noise come from the three-dimensional geological sedimentary model, which is manually intercepted layer by layer according to the model height to obtain multiple geological images with a large amount of noise.

[0075] The noise-free images are obtained by manually patching the geological images using drawing tools, and their number is extremely small.

[0076] As a result, the number of image pairs with matching river channels in the two types of images is very small, making them unusable for model training. Therefore, image segmentation methods are used to expand the image pair dataset.

[0077] First, an image preprocessing module is constructed to obtain image pairs corresponding to the river morphology, and then image pairs with a size of 128×128 are obtained by cutting them at different intervals.

[0078] Furthermore, in an embodiment of the present invention, the manually captured image pairs have black borders of varying widths, which may cause the river channel morphology of the cut image pairs to be inconsistent, and the different image sizes may also lead to inconsistent river channels. Therefore, an image preprocessing module is constructed to delete the black borders and redefine the size of the image.

[0079] Since the black pixel value is small and close to 0, a threshold is used to compare the pixel values ​​in the image one by one. If the pixel value is suddenly greater than the threshold, this position is the initial position without black edges in the image; if the pixel value is suddenly less than the threshold, this position is the end of the image without black edges. The present invention records the start and end of the horizontal and vertical black edges in the two images, and then cuts according to the obtained four positions to obtain a noise image without black edges and a noise-free image. The process can be expressed as:

[0080]

[0081]

[0082] Among them, flag is initialized to False, th=25 is the threshold, hMax[i] and hMax[j] are the maximum pixel values ​​of each column in the image; wMax[i] and wMax[j] are the maximum pixel values ​​of each row; h and w are the initial or end positions of the image without black edges that need to be recorded.

[0083] Furthermore, the images after black edge removal are adjusted to the same size to ensure the correspondence of the river channels in the image pairs before cutting. The image pairs of about 900×600 are repeatedly cut according to γ=(16,32,64) to obtain multiple 128×128 image pairs corresponding to the river channels.

[0084] Flipping is used to expand the dataset to ensure that the model can learn a variety of river morphology. Since the model is used to learn the location of noise points and the connectivity of the river channel, and does not focus on the specific morphology of the river channel, image flipping and cutting will not affect the model training. Finally, 784 noise-non-noise image pairs with different river channel morphology were obtained.

[0085] A noise detection model based on a multi-scale contrast training strategy is constructed based on the noise-noise image pair dataset:

[0086] See also Figure 2 , the model passes through the noisy image X noise With the noise-free image X good The contrast training between the two makes the generator gradually learn X noise Where is X good Different, thus capturing X noise The distribution of medium noise.

[0087] The generator in the model is based on the convolutional autoencoder for the input noise image Xnoise Encoding and decoding to obtain the reconstruction of the noisy image. However, a single autoencoder network cannot ensure that the information obtained is a useful feature representation, and it sometimes causes poor quality of the generated image due to the intervention of other noises. Therefore, the convolutional denoising autoencoder DAE is used to construct the generator to avoid the interference of noise on the generation process.

[0088] The original generator enhances the generation capability by minimizing the original geological image and the reconstructed image. This generator minimizes the reconstructed noisy image and the noise-free image to enhance the model's ability to capture noise distribution. The embodiment of the present invention transforms the function of the generator to make it more suitable for the detection of geological noise.

[0089] In order to better assist the generator in learning, the discriminator also changes from the generation / reconstruction discrimination to the presence / absence of noise discrimination. The discriminator can be regarded as an encoding process, and it has a similar CNN network structure to the encoder e1. to the discriminator, replacing Thus, the probability of the presence or absence of noise is obtained. The generation and discrimination process can be expressed as:

[0090]

[0091] Among them, f e1 (), f d () represents the encoding process and decoding process. D () is the discrimination process. θ e1 ,θ d ,θ D Represents the parameters in the training process. gaussian() represents the addition of noise. Z1 represents the vector obtained by encoder e1. Represents the noisy output of the discriminator and NoiselessX good The characteristic vector of Represents a noisy image with noise added.

[0092] The multi-scale contrast training strategy is to train the model from two perspectives: image space and feature space, so that it can better learn the distribution of complex noise points and discontinuities in the river channel. When training the generator, the model first minimizes the noise-free image X good Reconstruction of noisy images Make the generator according to X good Learn X noise Where the noise is.

[0093] Next X good and images The contrast loss in the latent space is used to strengthen the generator’s attention to the difference between the two from the feature level. Encoder e2 with the same structure as encoder e1 is used to extract the noise-free image X good The encoding feature Z2 is obtained by comparing Z2 with the vector Z1 obtained by encoder e1 to close the encoding process of the noise point and the noise-free point distribution, which helps the model to obtain more comprehensive noise detection results. The process can be expressed as:

[0094]

[0095] Loss z (Z1,Z2)=||Z1-Z2||1 (2)

[0096] Among them, Loss img and Loss z They are the contrast losses at the image level and feature level respectively;

[0097] The discriminator is trained using feature-level contrast loss. The features of the last layer of the discriminator and are fed into the discriminator, which come from the reconstructed noisy images and the noise-free image X good The embodiment of the present invention closes the discriminant feature distance from 0 and minimize the feature The distance from 1, the process is expressed as:

[0098]

[0099] Among them, Loss D Represents the overall loss of the model discriminator;

[0100] The maximum-minimum game method is used to train the noise detection model. When the generator is fixed to train the discriminator, the discriminator can accurately determine whether there is noise or not. When the discriminator is fixed to train the generator, the discriminator believes that the noise obtained from the model is Without noise, according to the proposed multi-scale contrast training strategy, the generator loss Loss G can be obtained. The overall loss can be expressed as:

[0101]

[0102] Among them, α, β, ρ are weight parameters, β = 50, α = ρ = 1. loss GAN is the original GAN ​​generator adversarial loss, which represents the minimum The distance from 1 makes the discriminator unable to distinguish.

[0103] Step 2: Build a pixel-level denoising model, remove some noise points in the original geological image based on the pixel-level denoising model, and obtain an image after removing some noise points;

[0104] The geological image is first subjected to a pixel-level denoising method to circle the non-river area and then fill the non-river area, so that the noise points in the area are modified to the corresponding correct pixel values, achieving the purpose of preliminary denoising.

[0105] Subsequently, according to the position of the noise point in the noise point detection result, the embodiment of the present invention can perform pixel value filling at the corresponding position in the noise point image to remove the noise point.

[0106] The cyclic denoising method is to perform model detection on the previous denoising result multiple times to obtain the noise point detection result, and then modify the pixel value at the corresponding position according to the position of the noise point to obtain a new denoising result.

[0107] The denoising results show that the river channel has good connectivity and integrity, the non-river area is clean and tidy, and most of the red noise points are basically removed, which fully proves that the two denoising methods can effectively remove noise points in geological images and have good denoising capabilities.

[0108] Specifically include:

[0109] See also Figure 3 , based on the idea of ​​contour detection, a pixel-level denoising method is constructed to preliminarily remove some noise points in the original geological image. The pixel-level denoising module is designed based on the idea of ​​contour detection in the field of natural image processing. It removes some noise points by circling the contours of non-river channel areas in the noise-filled image.

[0110] First, the geological image is binarized to obtain a black and white image to reduce the change in color depth and facilitate subsequent processing. Then the Gaussian filter is used to eliminate the interference of Gaussian noise to smooth the image. Gaussian filtering is a process of weighted averaging the entire image. Specifically, a 5×5 Gaussian kernel is used to scan the image horizontally and vertically. Then the weighted average of the pixels in the neighborhood determined by the Gaussian kernel replaces the central pixel value, and finally the Gaussian filtered image is obtained. Then the Flood Fill algorithm is used to mark the important areas in the river channel by darkening the color. The algorithm automatically selects the area connected to the seed pixel and fills the area color with the specified color, which has a good marking function for sensitive areas. The seed pixel value can be set to 150, and the newly assigned pixel value of the flood area can be set to (255,255,255), which is about 2-3 pixel values ​​different before and after. The color of the noise points and the discontinuities in the river channel in the flooded image is darker.

[0111] After flooding, the opening and closing operations are used to determine the physical form of the non-river area. The opening operation is used to discard the independent small red blocks inside the blue area and only select the overall coherent areas. The closing operation can fill the narrow blue pores inside the red area to ensure the connectivity of the large red area. After using the flood filling algorithm and the opening and closing algorithm, the embodiment of the present invention can basically determine the outline of the blue area. Finally, the non-river area is circled. However, some small areas are circled inside the river, which is an erroneous mark. In order to avoid errors as much as possible, the embodiment of the present invention sets an area threshold A=2000. Only areas with an area greater than the threshold will be circled. The embodiment of the present invention fills the circled area with color to obtain a binary image X with the noise points in the blue area removed. pixel .

[0112] Step 3: Obtain the generated result of the image after removing some noise through the noise detection model, compare the generated result with the image after removing some noise to obtain the noise position, remove the remaining noise based on the obtained noise position, and obtain the denoised and repaired image

[0113] The noise detection model is trained. After the model training is completed, the embodiment of the present invention inputs the entire noise image into the model to obtain the noise image generation result, which is biased towards the noise-free image river morphology corresponding to the input image. Due to the limitations of the network architecture, the quality of the generated image is limited, so the generation result is not used as a denoising result. The embodiment of the present invention subtracts the generation result from the input noise image to obtain the discontinuities and the locations of the noise in the input image, that is, the noise detection result. The noise detection result shows the location of most of the noise points, which fully proves the effectiveness of the detection model.

[0114] The specific steps include:

[0115] A cyclic denoising method is constructed based on the trained model to completely remove the remaining noise points. Denoising can be regarded as a process of modifying pixel values. Specifically, the blue noise points inside the river channel can be modified to red to ensure the integrity and connectivity of the river channel; the red noise points outside the river channel can be modified to blue to ensure the integrity of the non-river channel area. However, one modification operation cannot complete two denoising processes at the same time. Therefore, the cyclic denoising method cyclically modifies the pixel values ​​to completely denoise, and it calls the model multiple times to prevent noise points from being missed.

[0116] Given a noisy image X noise First, the pixel-level image X is obtained by pixel-level denoising. pixel Then set X pixel Input into the model to detect the remaining noise points, thus obtaining the noise detection image X detect The embodiment of the present invention is based on X detectFind the corresponding noise points on the noise image and fill them with red / blue colors. Then we get the first noise repair image X. result1 Then, the embodiment of the present invention will result1 Put it into the model and detect the noise again to get the second detection result X detect2 , and then according to X detect2 X result1 After C similar restoration processes, the embodiment of the present invention obtains a result image X with a complete river channel, good connectivity and no noise. result . The process can be expressed as:

[0117]

[0118] Where F1() and F c () represent the first and cth cycle denoising respectively. Model() is the process of model detecting noise points. X resultc is the result obtained after the cth cycle. is the result obtained in the c-1th cycle.

[0119] Due to the single architecture of the network, the quality of the generated image obtained by the model is poor. Therefore, the generated image cannot be directly regarded as the denoising result. Therefore, the embodiment of the present invention subtracts the generated image from the geological image of the input model on the three channels (R, G, B) to obtain the detection result X with the noise point position detect . The process can be expressed as:

[0120] X detect =f threshold (MAX(Abs R ,Abs G ,Abs B ))

[0121]

[0122] Among them, Abs R , Abs G , Abs B represents the subtraction operation between the geological image and the generated result on (R, G, B), f threshold The function is to binarize the noise detection result. The best effect is when threshold=100.

[0123] See also Figure 4 , according to the model, most of the noise points can be detected. A geological image and its corresponding detection results are sliced ​​to obtain Figure 4The four sets of images on the right show that the model can detect the noise points in the non-river area that are not detected by the pixel-level denoising method. The model can accurately capture the horizontal and vertical bars inside and outside the river. At the same time, the model can accurately detect the discontinuities between the river channels, thus ensuring the connectivity and integrity of the river channels.

[0124] Since the red river channels account for a large proportion in geological images, the embodiment of the present invention can repeat the process of filling with red for multiple times in the method to ensure that the noise points inside the red river channels in the obtained results are completely removed. Since the pixel-level denoising has removed most of the noise points in the blue area, the embodiment of the present invention only chooses to modify the noise points to blue at one time in the middle. The method disclosed in the embodiment of the present invention can use the model for multiple detections to minimize the errors caused by model omissions, thereby ensuring the comprehensiveness of denoising.

[0125] See also Figure 5 , the result map obtained by cyclic denoising shows good river connectivity and integrity. And Figure 5 The detection results obtained by calling the model 5 times on the right show that the number of noise points decreases as the number of times increases. This also indirectly proves that the cyclic denoising method is indeed repairing and removing noise points.

[0126] See also Figure 6 , which are the denoising results of different river channels obtained by pixel-level denoising method. Figure 6 As shown in Figure 2, the noise points in the large non-river area can be removed. Although the denoising effects of different rivers are different, this method can always ensure the removal of some noise points, thereby alleviating the pressure of cyclic denoising to remove noise points.

[0127] See also Figure 7 Compared with the unsupervised method, this method is significantly better than the unsupervised method in five categories of similarity, which fully proves the superiority of the contrast training between noise and noise-free images. Compared with the AE and GAN methods under the same dataset, this method is generally better than the other two methods, which shows that the joint contrast of image space and feature space is conducive to the detection of noise.

[0128] See also Figure 8 , is a comparison between the noise point detection result and the denoising result obtained by the unsupervised automatic denoising method based on visual image features and the result obtained by the present invention. Figure 8As shown, according to the noise detection results in the third and fourth rows, the noise detection model of the present invention can detect more noise points, it can well obtain the discontinuities between river channels, and the noise detection number density is significantly better than the unsupervised method. According to the denoising results in the first and second rows, the model and method proposed in the present invention obviously better ensure the integrity and connectivity of the river channel. Although the unsupervised method effectively removes some noise points, the river channel connectivity in its denoising results is not good. The method of the present invention effectively ensures the neatness and integrity of non-river areas. The embodiment of the present invention constructs a deep learning method for noise detection and removal.

[0129] In the noisy image obtained by slicing the model, the noise can be regarded as an "abnormal point", and noise detection can be regarded as abnormal detection. Since the noise distribution is disorderly and difficult to mark, it is impossible to make a label image of the noise, which means that many supervised abnormality detection methods are difficult to apply to denoising. Therefore, the reconstruction-based abnormality detection in the embodiment of the present invention is studied for noise detection, and the generative adversarial network GAN is used to reconstruct the image. Specifically, the model learns the good river channel features of the noise-free image by continuously reconstructing the image, thereby repairing the input noise image.

[0130] In order to better determine the discontinuities in the river channel and avoid the phenomenon of poor connectivity in the river channel, an embodiment of the present invention simultaneously trains noisy images and noise-free images. By reconstructing the noisy images and comparing them with the noise-free images, the model can better learn the noise distribution and the discontinuities in the river channel, and compare them from both the image space and the feature space to achieve better detection effects.

[0131] In addition, due to the poor quality of image reconstruction caused by the single network architecture, the reconstruction cannot be directly used as the final result image. The embodiment of the present invention uses the difference between the noisy image and the reconstruction as the noise detection image, and then uses recursion to continuously fill the noisy image to generate a clear result image of the complete river channel.

[0132] The method disclosed in the embodiment of the present invention can obtain a data set of noisy-noise-free image pairing by some image expansion methods, avoiding complex work such as noise mark, and saving human resources. The present invention automatically removes noise after automatic noise detection by computer, thereby saving a lot of human resources. The denoising result obtained shows the river morphology with good integrity and connectivity, which provides help for oil and gas reservoir prediction.

[0133] The embodiment of the present invention discloses a geological model image denoising system:

[0134] It includes a noise detection model building module, a first noise removal module and a second noise removal module;

[0135] A noise detection model building module is used to build a paired dataset of noisy images and noise-free images, and to build a noise detection model based on the paired dataset;

[0136] The first noise removal module is used to construct a pixel-level denoising model, remove some noise points in the original geological image based on the pixel-level denoising model, and obtain an image after removing some noise points;

[0137] The second noise removal module is used to obtain the generation result of the image after removing part of the noise through the noise detection model, compare the generation result with the image after removing part of the noise to obtain the noise position, remove the remaining noise based on the obtained noise position, and obtain the denoised and repaired image.

[0138] The present invention learns the difference between noise and noise-free images through feature-level and image-level comparison. A dataset of noise-free image pairs is constructed for comparative training. In addition, the pixel-level denoising method preliminarily removes some noise points by delineating the contours of non-river areas, which reduces the pressure of cyclic denoising. The cyclic denoising method cyclically fills the noise image according to the detected noise points to obtain a complete denoising result for the river channel. The result graph of the detection shows that the model can well detect the discontinuities between noise points and the river channel. The denoising result shows the good connectivity of the river channel, making it more in line with geological laws.

[0139] A schematic diagram of a terminal device provided in an embodiment of the present invention. The terminal device of this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps in the above-mentioned method embodiments are implemented. Alternatively, when the processor executes the computer program, the functions of the modules / units in the above-mentioned device embodiments are implemented.

[0140] The computer program may be divided into one or more modules / units, and the one or more modules / units are stored in the memory and executed by the processor to accomplish the present invention.

[0141] The terminal device may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0142] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0143] The memory may be used to store the computer programs and / or modules, and the processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory.

[0144] If the module / unit integrated in the terminal device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0145] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for denoising a geological model image, characterized in that: The following steps are involved: Construct a paired dataset of noisy images and noise-free images, and build a noise detection model based on the paired dataset; Construct a pixel-level denoising model, remove some noise points in the original geological image based on the pixel-level denoising model, and obtain an image after removing some noise points; The noise detection model is used to obtain the generated result of the image after removing some noise points. The generated result is compared with the image after removing some noise points to obtain the noise point position. The remaining noise points are removed based on the obtained noise point position to obtain the denoised and repaired image.

2. A geological model image denoising method according to claim 1, characterized in that: The step of constructing a paired dataset of noisy images and noise-free images comprises the following steps: Acquire geological images with noise and geological images without noise, and preprocess the geological images with noise and geological images without noise so that each pair of image channels correspond to each other, and obtain image pairs corresponding to the channel morphology; By cutting with different intervals γ, we get the image pair dataset.

3. A geological model image denoising method according to claim 1, characterized in that: It also includes training the noise detection model: The model minimizes the noise-free image X good Reconstruction of noisy images Make the generator according to X good Learn X noise The location of the medium noise point; Next X good and images The contrast loss in the latent space is used to strengthen the generator's attention to the difference between the two from the feature level; Encoder e2 with the same structure as encoder e1 is used to extract the noise-free image X good The encoding feature Z2 is compared with the vector Z1 obtained by encoder e1 to bring the encoding process of the noise point and the noise-free point distribution closer: Loss z (Z1,Z2)=||Z1-Z2||1 (2) Among them, Loss img and Loss z They are the contrast losses at the image level and feature level respectively; Generator Loss G for: Among them, α, β, ρ are weight parameters, β = 50, α = ρ = 1; loss GAN is the original GAN ​​generator adversarial loss, which represents the minimum The distance from 1 makes the discriminator unable to distinguish; The discriminator is trained by feature-level contrast loss, and the features of the last layer of the discriminator are and Used for loss judgment, to bring the discriminant features closer distance from 0 and minimize the feature Distance from 1: Among them, Loss D represents the overall loss of the model discriminator, Represents the noisy image output by the discriminator The eigenvector of Represents the noise-free image X output by the discriminator good The feature vector of .

4. A geological model image denoising method according to claim 1, characterized in that: The method of removing some noise points existing in the original geological image based on the pixel-level denoising model includes: Binarization processing is performed on the geological image to be denoised to obtain a black and white image; The obtained black and white image is processed by a Gaussian filter to obtain a weighted average value of pixels in the neighborhood, and the weighted average value of pixels in the neighborhood is used to replace the central pixel value to obtain a Gaussian filtered image; Based on the Gaussian filtered image, the Flood Fill algorithm is used to highlight the important areas in the river. Based on the image obtained after Flood Fill, the non-river channel area in the geological image is determined and circled through opening and closing operations, the circled area is filled with color, and the noise points in the area are modified to the corresponding correct pixel values ​​to achieve preliminary denoising.

5. A method for denoising a geological model image according to claim 4, characterized in that: The Flood Fill algorithm is used to color-enhance important areas in the river, including: The seed pixel value can be set to 150, and the newly assigned pixel value of the flood area can be set to (255,255,255). The difference between the pixel values ​​before and after is about 2-3. The color of the noise points and the discontinuities in the river channel in the flooded image is darker. The opening and closing operations to determine the non-river channel area in the geological image include: Use the opening operation to discard the independent small red blocks inside the blue area and only select the overall coherent areas; The closing operation is used to fill the blue gaps inside the red area to ensure the connectivity of the large red area, and then determine the outline of the blue area, that is, to delineate the non-river area.

6. A geological model image denoising method according to claim 1, characterized in that: The process of comparing the image after removing some noise points with the corresponding model generation result includes: Given a noisy image X noise , after pixel-level denoising, we get the pixel-level image X pixel ; X pixel Input into the noise detection model to detect the remaining noise points and obtain the noise detection image X detect According to X detect Find the corresponding noise points on the noise image and fill them with red or blue colors to obtain the noise repair image X. result1 .

7. A geological model image denoising method according to claim 6, characterized in that: The noise detection image X detect The acquisition process is: Subtract the generated image from the geological image input to the model on the three channels (R, G, B) to obtain the detection result X of the location with noise detect : X detect =f threshold (MAX(Abs R ,Abs G ,Abs B )) Among them, Abs R , Abs G , Abs B represents the subtraction operation between the geological image and the generated result on (R, G, B), f threshold The function is to binarize the result of noise detection.

8. A geological model image denoising system, characterized in that: It includes a noise detection model building module, a first noise removal module and a second noise removal module; A noise detection model building module is used to build a paired dataset of noisy images and noise-free images, and to build a noise detection model based on the paired dataset; The first noise removal module is used to construct a pixel-level denoising model, remove some noise points in the original geological image based on the pixel-level denoising model, and obtain an image after removing some noise points; The second noise removal module obtains the generated result of the image after removing part of the noise through the noise detection model, compares the generated result with the image after removing part of the noise to obtain the noise position, removes the remaining noise based on the obtained noise position, and obtains the denoised and repaired image.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images