Geological texture synthesis method based on GAN network

Abstract

The invention discloses a geologic texture synthesis method based on a GAN network, and the method comprises the steps: obtaining an original geologic texture picture, and carrying out the screening and preprocessing of the original geologic texture picture, and obtaining a geologic texture data set; calculating homogeneity, repeatability and directivity values of each picture sample in the geological texture data set to obtain a sample index range, and performing sample screening on the geological texture data set based on the sample index range to obtain a typical sample data set; and constructing a geologic texture synthesis network model based on a generative adversarial network, and inputting the typical sample data set into the geologic texture synthesis network model to obtain a synthesized geologic texture picture. Through accurate sample screening and an efficient generation algorithm, real geological texture mapping resources can be provided for the field of three-dimensional geological modeling, and researchers are helped to better understand the geological structure and form.

Description

Technical Field

[0001] The present invention relates to the field of three-dimensional geological science and technology, and in particular to a geological texture synthesis method based on a GAN network. Background Art

[0002] In the field of geology, 3D geological models have the ability to express information such as complex structures, numerous components and irregular shapes of geological units, and are therefore widely used in mineral resource exploration, mine planning and management, urban underground space resource management, road and bridge construction, and other fields. Through visualization technology, 3D geological models can render 3D geological entities with good visual effects, and clearly show the geological spatial structure and change patterns reflected by the model. In the process of 3D model rendering, texturing is a fast and effective means to increase the realism of the model, and texture mapping is an important rendering resource.

[0003] At present, geological three-dimensional models generally use color or color superimposed simple lines as the texture map of the model for texture mapping. When mapping, on the one hand, because geological texture is usually affected by factors such as noise, lighting conditions, shooting angle and object surface shape, the area of ​​texture extracted directly from the geological texture image obtained by shooting may not be enough to cover the entire three-dimensional model surface. On the other hand, due to the unevenness of geological texture, when drawing model texture, directly drawing geological texture map on the large surface of the model can easily lead to distortion of the drawn model texture and loss of the model's realism; while using texture wrap technology will cause surface texture seams and distortion problems. It can be seen that high-quality geological texture map resources are an important factor in the realism of three-dimensional geological models. It is of practical significance to study how to obtain high-quality geological texture map resources. It can not only assist the visualization of geological information and provide intuitive visual presentation for geological research, but also help geologists better understand geological structure and morphology, and have a positive impact on the field of geological engineering.

[0004] At present, there are two ways to obtain high-quality and stable texture map resources. The first is that digital artists use software tools such as Substance Painter to create textures. The second is an automatic method based on a special algorithm, namely texture synthesis. Texture synthesis refers to the process of generating textures of any size. Texture synthesis is currently used in many fields, such as image restoration and style transfer, providing theoretical and technical support for the development of many fields. Texture synthesis is divided into procedural texture and sample-based texture synthesis. Procedural texture can generate many approximately regular textures, but in actual use, a slight change in the value of one or more parameters may result in two texture images with completely different appearances, which increases the uncertainty of texture generation and makes it difficult to predict the appearance of the texture based on the input parameter value. Compared with procedural texture, the sample-based method has the ability to reproduce textures with similar appearance to the input sample and avoids cumbersome parameter design. The traditional sample-based texture synthesis method is developed based on the assumption of Markov random field. It constructs a larger texture by mechanically copying the local sample, ignoring the global visual characteristics of the texture.

[0005] Therefore, it is necessary to propose a geological texture synthesis method based on the GAN network, which can generate more realistic textures that conform to global visual consistency, avoid the seam problem of local texture splicing in traditional methods, improve the quality and stability of geological texture synthesis, and generate high-quality texture maps suitable for geological three-dimensional models. Summary of the invention

[0006] In view of this, the present invention provides a geological texture synthesis method based on a GAN network to solve the technical problems of large uncertainty in texture generation in the current geological texture synthesis method and the neglect of the global visual features of the texture, resulting in inaccurate and unrealistic synthesized geological texture images.

[0007] The present invention provides a geological texture synthesis method based on a GAN network, comprising:

[0008] Acquire an original geological texture image, screen and preprocess the original geological texture image to obtain a geological texture data set;

[0009] Calculating the homogeneity, repeatability and directionality values ​​of each image sample in the geological texture dataset to obtain a sample index range, and performing sample screening on the geological texture dataset based on the sample index range to obtain a typical sample dataset;

[0010] A geological texture synthesis network model based on a generative adversarial network is constructed, and the typical sample data set is input into the geological texture synthesis network model to obtain a synthetic geological texture image.

[0011] Furthermore, the calculating of the homogeneity value of each image sample includes:

[0012] Expanding the RGB channel values ​​of the image sample as the visual vocabulary of the image sample;

[0013] Clustering the visual vocabulary using a K-means algorithm, merging words with similar meanings in the visual vocabulary, and generating a visual dictionary;

[0014] According to the visual dictionary, the image samples are sampled in local areas of the same size, the underlying features of each local area are extracted and encoded, the labels of new data are determined according to the labels of surrounding neighboring data using a k-nearest neighbor algorithm, and the underlying features are quantized to the most similar visual vocabulary to construct a visual vocabulary vector of the local area;

[0015] The Euclidean distance calculation is performed on the visual vocabulary vectors of any two local areas in the image sample, the calculation is repeated for a preset number of times, and the average value is used as the homogeneity value of the sample image.

[0016] Furthermore, the calculation of the repeatability value of each image sample includes:

[0017] Perform correlation function calculation on the image sample and itself to obtain an autocorrelation matrix;

[0018] Clipping the border of the autocorrelation matrix, and normalizing the clipped autocorrelation matrix to obtain a normalized autocorrelation matrix;

[0019] A plurality of local areas in the image sample are randomly selected, a maximum value, a minimum value and a standard deviation of each local area are calculated, and the calculation results of the plurality of local areas are averaged to obtain a repeatability value of the image sample.

[0020] Furthermore, the calculation formula of the repeatability value of the image sample is:

[0021]

[0022] Among them, Max(Rt) and Min(Rt) represent the maximum and minimum values ​​of the random region Rt, respectively, σ(Min(Rt)) is the standard deviation of the minimum value, and T represents the number of randomly selected regions.

[0023] Further, the calculating of the directionality value of each image sample includes:

[0024] Calculate the magnitude and angle of the gradient vector of each pixel of the image sample, and obtain the valid pixel point based on the preset gradient magnitude threshold;

[0025] Divide the angles of the effective pixel points into n types of directions, and obtain a direction histogram of the sample image by statistics;

[0026] The histogram of any category direction is expressed as:

[0027]

[0028] Among them, N θ (φ) represents the number of pixels with direction category φ in category n;

[0029] The directivity value is calculated as:

[0030]

[0031] in, Indicates the direction Histogram of n p represents the number of peaks in the histogram, p represents the histogram peak, and w p Indicates the direction category range, Indicates the class of the peak in the histogram.

[0032] Furthermore, the construction of a geological texture synthesis network model based on a generative adversarial network includes:

[0033] Add a multi-scale convolution layer to the adversarial network to extract multi-scale features of the image and output a multi-scale feature map;

[0034] A channel attention mechanism is added on the basis of multi-scale feature extraction, the multi-scale feature map is weighted and adjusted to obtain multi-scale weighted features, and the channel weights are redistributed according to the importance of each channel;

[0035] Introducing shared source residual blocks, each residual block uses the same input source feature map and uses shared convolution weights to fuse features of different scales to learn multi-scale feature information;

[0036] A content-aware upsampling module is introduced to combine the features of the input image to restore the detailed information of the geological texture.

[0037] Furthermore, the generator of the geological texture synthesis network model includes a multi-branch convolutional layer and a multi-scale attention module connected to the multi-branch convolutional layer;

[0038] The multi-branch convolution layer is used to extract multi-scale features from the image. Each branch performs convolution calculation using grouped convolution to output a multi-scale feature map. The number of groups is dynamically adjusted according to the size of the convolution kernel.

[0039] The multi-scale attention module is used to adjust the inter-channel information of the multi-scale feature map, obtain the channel attention weights of each scale, and integrate the multi-scale attention in a cascade manner to realize the interaction of attention information at multiple scales.

[0040] Furthermore, the content-aware upsampling module includes a sampling unit, a channel reduction unit, and a feature fusion unit connected in sequence;

[0041] The sampling unit is used to perform 2-fold upsampling through CARAFE to obtain an upsampled feature map;

[0042] The channel reduction unit uses Conv3×3-BN-LeakyReLU to reduce the channel;

[0043] The feature fusion unit is used to perform an addition operation on the feature map after channel reduction and the original feature map to obtain a feature map that restores geological texture detail information.

[0044] Furthermore, the screening of the original geological texture images includes:

[0045] Remove the images covered by interfering textures and keep the images with clear rock textures;

[0046] Remove overexposed or shadowed images;

[0047] Remove images where the angle between the shooting angle and the geological texture is greater than the inclination threshold.

[0048] Furthermore, the preprocessing includes image cropping and image culling;

[0049] The image cropping process is used to crop the geological texture image into a uniform size;

[0050] The image culling process is used to splice all geological texture images in the vertical and horizontal directions, and to cull texture samples that can be seamlessly spliced.

[0051] Compared with the prior art, this method firstly screens and preprocesses the original geological texture images, effectively eliminating samples with poor quality or not meeting the requirements, which not only reduces the interference of invalid data, but also improves the accuracy of subsequent processing; secondly, by calculating the homogeneity, repeatability and directionality values, the sample quality is evaluated from different dimensions, and the sample selection is further optimized, which is conducive to improving the learning effect of the model, generating synthesis results that are more in line with the characteristics of real geological textures, avoiding excessive reliance on a certain type of texture, and ensuring that the training data set covers more texture diversity; finally, a geological texture synthesis model is constructed based on a generative adversarial network (GAN), and the GAN network is improved, and the training data set is combined to synthesize high-quality, diversified and realistic geological texture images, and finally obtain high-quality geological texture map resources. The present invention screens and preprocesses the original images, and uses quantitative indicators to calculate the texture content of the collected geological textures, further optimizes the data set, and outputs high-quality geological texture images by constructing a geological texture synthesis model, which solves the problems of poor quality and insufficient diversity in traditional texture synthesis methods. Through accurate sample screening and efficient generation algorithms, it can provide real geological texture materials for the field of three-dimensional geological modeling, and also improve the efficiency of texture synthesis. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic diagram of the process of the geological texture synthesis method based on the GAN network provided by the present invention;

[0053] Figure 2 A schematic diagram of the seamless splicing test of samples provided by the present invention;

[0054] Figure 3 A sample data acquisition flow chart provided by the present invention;

[0055] Figure 4 A schematic diagram of the texture image homogeneity calculation process provided by the present invention;

[0056] Figure 5 A schematic diagram of the texture image directionality calculation process provided by the present invention;

[0057] FIG6( a ) is a schematic diagram of a typical sample of geological texture provided by the present invention;

[0058] FIG6( b ) is a schematic diagram of another typical sample of geological texture provided by the present invention;

[0059] Figure 7 A schematic diagram of the multi-scale convolution structure provided by the present invention;

[0060] Figure 8 A flowchart of the implementation of the geological texture synthesis method provided by the present invention;

[0061] Fig. 9 This is the geological feature texture synthesis result map provided by the present invention. DETAILED DESCRIPTION

[0062] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0063] The present invention provides a geological texture synthesis method based on a GAN network, which is described in detail below.

[0064] Example 1

[0065] See also Figure 1 This embodiment provides a geological texture synthesis method based on a GAN network, including:

[0066] Step S101: obtaining an original geological texture image, screening and preprocessing the original geological texture image to obtain a geological texture data set;

[0067] Step S102: Calculate the homogeneity, repeatability and directionality values ​​of each image sample in the geological texture dataset to obtain a sample index range, and perform sample screening on the geological texture dataset based on the sample index range to obtain a typical sample dataset;

[0068] Step S103: constructing a geological texture synthesis network model based on a generative adversarial network, inputting the typical sample data set into the geological texture synthesis network model, and obtaining a synthetic geological texture image.

[0069] The method of this embodiment firstly screens and preprocesses the original geological texture images, effectively eliminating samples with poor quality or not meeting the requirements, which not only reduces the interference of invalid data, but also improves the accuracy of subsequent processing; secondly, by calculating the homogeneity, repeatability and directionality values, the sample quality is evaluated from different dimensions, and the sample selection is further optimized, which is conducive to improving the learning effect of the model, generating synthesis results that are more in line with the characteristics of real geological textures, avoiding excessive reliance on a certain type of texture, and ensuring that the training data set covers more texture diversity. Finally, a geological texture synthesis model is constructed based on a generative adversarial network (GAN), and the GAN network is improved, and high-quality, diversified, and realistic geological texture images are synthesized in combination with the training data set, and finally a high-quality geological texture map resource is obtained. This embodiment screens and preprocesses the original images, and uses quantitative indicators to calculate the texture content of the collected geological textures, further optimizes the data set, and constructs a geological texture synthesis model to output high-quality geological texture images, which solves the problems of poor quality and insufficient diversity in traditional texture synthesis methods. Through accurate sample screening and efficient generation algorithms, it can provide real geological texture materials for the field of three-dimensional geological modeling, and also improve the efficiency of texture synthesis.

[0070] Example 2

[0071] In this embodiment, firstly, "ROCK" classified pictures from texture websites such as Texture website, Texture Ninja website, Hippopx website and AlexStrekeisen and pictures related to the keyword "rock texture" from photography websites (ie, original geological texture pictures) are collected.

[0072] The original geological texture pictures were screened according to the following criteria: 1. The rock texture in the picture is clear and not completely covered by other textures (such as moss, grass, etc.); 2. The picture has uniform lighting without overexposure or shadows caused by light; 3. The shooting angle is as close to the geological texture as possible, and there is no obvious tilt in the angle of view, thus obtaining 166 original pictures.

[0073] Then, the filtered images were preprocessed: 1. The images were randomly cropped to a size of 300×300; 2. All geological texture images were spliced ​​in the vertical and horizontal directions, and texture samples that could be seamlessly spliced ​​were eliminated, such as Figure 2 As shown, the geological texture images that can be seamlessly spliced ​​are removed.

[0074] Finally, 110 images were retained from the 166 original images to form the geological texture data set. The sample data acquisition flow chart of this embodiment is as follows: Figure 3 shown.

[0075] Furthermore, this embodiment uses three quantitative texture content indicators - homogeneity, repeatability and directionality to examine the texture content of geological textures, analyze the characteristics of geological textures, and screen out typical geological texture samples (low homogeneity, low repeatability, and obvious directionality) and common geological feature texture samples to verify the effect of the network of the present invention on geological texture synthesis.

[0076] (1) Calculating the homogeneity value of each image sample includes:

[0077] The RGB channel values ​​are expanded as the visual vocabulary of the image, and the obtained visual vocabulary is clustered using the K-means algorithm to map the features from high dimension to low dimension. The words with similar meanings in the image visual vocabulary are merged as the words in the visual dictionary to generate a visual dictionary.

[0078] According to the visual dictionary, the underlying features of two local areas of the same size randomly sampled from the image are encoded, and the k-nearest neighbor algorithm is used to quantify the local features of the local area to its most similar visual vocabulary. The k-nearest neighbor algorithm determines the label of the new data through the labels of the surrounding neighboring data, converts the underlying features of the local area into the vocabulary representation in the visual dictionary, and constructs the visual vocabulary vector of the local area. By counting the frequency of each visual word in the local area of ​​the image, a visual vocabulary frequency histogram of the local area of ​​the image is obtained;

[0079] The Euclidean distance between any two local areas in the image is performed multiple times to calculate the similarity between the two vectors, and the average value is taken as the homogeneity of the image.

[0080] The above process is illustrated below using a specific example.

[0081] Suppose we have an image of rock texture (such as a picture of a marble), which contains different color and texture features. Each pixel of the image can be represented by a three-dimensional vector. The three values ​​correspond to the brightness of the RGB color channels. Each pixel has three color channel values ​​(R, G, B). We extract the RGB values ​​of all pixels in this image as the "visual vocabulary" of the image. That is, each pixel of the image is regarded as a vocabulary unit, and its color information (RGB value) is the feature of this unit.

[0082] The K-means algorithm is used to cluster these RGB values ​​(as features), with the goal of clustering pixels with similar RGB color features into the same cluster. Assuming we choose K=10, K-means will divide all pixel RGB values ​​into 10 clusters, with similar RGB values ​​in each cluster, representing different "visual vocabularies". In this way, the high-dimensional data of the RGB channel (each pixel consists of three channels, R, G, and B, and each channel is a continuous value) will be reduced to 10 cluster centers by K-means clustering, that is, the high-dimensional RGB features are mapped into low-dimensional "vocabulary". Each cluster center represents a specific visual feature (such as a certain color or texture in a rock), and each pixel can be represented by one of these vocabularies.

[0083] The k-nearest neighbor algorithm determines the closest visual vocabulary for a local area by comparing the similarities between different feature vectors. In this way, each local area in the image is assigned a visual vocabulary label.

[0084] For any two local regions in an image, their visual vocabulary vectors can be viewed as a high-dimensional vector. By calculating the Euclidean distance between these two vectors, we can measure the similarity of the two local regions in visual vocabulary. The smaller the Euclidean distance, the more similar the visual features of the two regions are, and vice versa.

[0085] like Figure 4 As shown, Figure 4 The specific steps of the numerical calculation of homogeneity are demonstrated.

[0086] (2) Calculating the repeatability value of each image sample includes:

[0087] Perform correlation function calculation on the image sample and itself to obtain an autocorrelation matrix;

[0088] The border of the autocorrelation matrix is ​​cropped because the border of the matrix does not overlap enough, so it needs to be cropped, and the cropped autocorrelation matrix is ​​normalized; wherein the normalization method is: dividing each element in the matrix by the number of pixels calculated;

[0089] A plurality of local areas in the image sample are randomly selected, the maximum value, the minimum value and the standard deviation of each local area are calculated, and the repeatability of the local area of ​​the image is calculated. To ensure accuracy, the calculation results of the plurality of local areas are averaged to obtain the repeatability value of the image sample.

[0090] The calculation formula of the repeatability value of the image sample is:

[0091]

[0092] Among them, Max(Rt) and Min(Rt) represent the maximum and minimum values ​​of the random region Rt, respectively, σ(Min(Rt)) is the standard deviation of the minimum value, and T represents the number of randomly selected regions.

[0093] (3) Calculating the directional value of each image sample includes:

[0094] Calculate the modulus and angle of the gradient vector of each pixel of the image sample, and obtain the valid pixel points (filter out the pixels with strong directionality) based on the preset gradient modulus threshold;

[0095] Specifically, the angles of the effective pixel points are divided into n types of directions, and a direction histogram of the sample image is obtained by statistics;

[0096] The histogram of any category direction is expressed as:

[0097]

[0098] Among them, N θ (φ) represents the number of pixels with direction category φ in category n;

[0099] The directivity value is calculated as:

[0100]

[0101] in, Indicates the direction Histogram of n p represents the number of peaks in the histogram, p represents the histogram peak, and w p Indicates the direction category range, Indicates the class of the peak in the histogram.

[0102] The more obvious the main directionality of the texture image is, the smaller the value will be, and vice versa.

[0103] The calculation process of the directional value is shown in the figure below: Figure 5 shown.

[0104] Finally, the independence of the indicators was verified based on the calculation of the Pearson coefficient and maximum likelihood coefficient of the three indicators of the texture data set. The homogeneity, repeatability and directionality values ​​of each geological texture in the data set were calculated one by one. The properties of geological textures were analyzed based on the statistical range of homogeneity, repeatability and directionality indicators of the geological texture data set. The selection criteria of the three indicators were determined according to the properties of geological textures, so that geological texture images with low homogeneity, low repeatability and low directionality (obvious main directionality) were selected as typical samples.

[0105] Typical samples are shown in Figure 6(a) and Figure 6(b). In this example, the geological texture image with a clear main direction and the texture image with low homogeneity are the same, that is, Figure 6(a). This image shows the trend of the geological texture, and the similarity between its local areas is very low. Figure 6(b) is a geological texture image with low repeatability, and it is difficult to extract repeated elements in the image.

[0106] Example 2

[0107] This embodiment designs a geological texture synthesis network model based on the generative adversarial network. Combining multi-scale convolution and channel attention, channel attention is introduced on the basis of multi-scale feature extraction, channels are redistributed according to their importance, and shared source residual blocks are combined so that each subsequent residual block can learn multi-scale information. A content-aware upsampling module is introduced, which can select appropriate features for flexible combination according to the input feature map, and can better restore the detail information of the texture image after extracting the features.

[0108] The construction of a geological texture synthesis network model based on a generative adversarial network includes:

[0109] Add a multi-scale convolution layer to the adversarial network to extract multi-scale features of the image and output a multi-scale feature map;

[0110] A channel attention mechanism is added on the basis of multi-scale feature extraction, the multi-scale feature map is weighted and adjusted to obtain multi-scale weighted features, and the channel weights are redistributed according to the importance of each channel;

[0111] Introducing shared source residual blocks, each residual block uses the same input source feature map and uses shared convolution weights to fuse features of different scales to learn multi-scale feature information;

[0112] A content-aware upsampling module is introduced to combine the features of the input image to restore the detailed information of the geological texture.

[0113] Specifically, Figure 7 As shown in the figure, multiple branch convolution layers are introduced into the generator of the generative adversarial network. Group convolution is applied to each branch of the multi-branch structure for convolution calculation. The network will dynamically adjust the size of the group according to the convolution kernel. For convolution kernels larger than 3, the relationship between the number of groups satisfies the following formula:

[0114]

[0115] Among them, K×K represents the convolution kernel size.

[0116] Second-order Channel Attention (SOCA) is introduced to adjust information between channels and filter out useless feature expressions in the network. The channel attention SOCA weight is combined with the multi-scale grouped convolution to form a multi-scale attention module. The input feature map is processed by the multi-scale grouped convolution module to extract spatial information of different scales from each channel feature map. The information of different scales is merged, and the multiple scale feature maps preprocessed by the multi-scale grouped convolution are input into the SOCA weight module to obtain the channel attention weights of each scale. The multi-scale attention is integrated in a cascade manner to realize the interaction of attention information of multiple scales.

[0117] A shared source residual block is introduced to fuse features from multiple convolutional branches, and skip connections are used to improve the stability and information transfer capability of the network.

[0118] The introduction of the content-aware upsampling module involves upsampling the input feature map by a factor of 2 through CARAFE, then reducing the channel through Conv3×3-BN-LeakyReLU, and then fusing it with the feature map from the downsampling layer by addition operation.

[0119] It should be noted that in addition to the structural improvements of the generator part, the discriminator is also adaptively optimized. For example, the discriminator also introduces a multi-scale discrimination mechanism to evaluate the authenticity of the generated image at different scales, introduces a channel attention mechanism, dynamically adjusts the importance of different channels in the process of judging the generated image, and strengthens the focus on important feature channels. In addition to the traditional adversarial loss (such as binary cross entropy loss), it can also be combined with content-aware loss, especially with the content-aware module in the generator (such as CARAFE). Through content-aware loss, the discriminator not only evaluates the authenticity of the image, but also evaluates the similarity between the image details and the original geological texture image. This method can help the generator generate images that are more detailed and close to real geological textures.

[0120] The complete process of texture synthesis of the original geological texture image by the geological texture synthesis network model of this embodiment is as follows: Figure 8 Each image after screening, preprocessing and quantitative analysis is input into the network for training to obtain a synthetic geological texture image. The synthetic result is shown in Fig. 9 Shown are typical geological textures and their texture synthesis images.

[0121] from Fig. 9 It can be seen that the texture images synthesized by the method of the present invention have obvious advantages in detail restoration, can synthesize more refined and diverse texture images, and can provide high-quality geological texture map resources for the construction of three-dimensional geological models.

[0122] In summary, the geological texture synthesis method based on the GAN network provided in this embodiment first collects and screens geological texture images through the network, preprocesses the acquired geological texture images, and extracts geological feature samples; calculates the homogeneity, repeatability and directionality values ​​of each geological texture in the data set one by one. Based on the statistical homogeneity, repeatability and directionality index range of the geological texture data set, geological texture images with low homogeneity, low repeatability and low directionality are selected as typical samples; a geological texture synthesis network model based on a generative adversarial network is constructed, combining multi-scale convolution and channel attention, introducing channel attention on the basis of multi-scale feature extraction, reallocating channels according to the importance of channels, and combining shared source residual blocks to introduce content-aware upsampling modules. Based on the designed network model, images are input one by one for training to obtain synthesized geological texture images.

[0123] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A geological texture synthesis method based on GAN network, characterized in that: include: Acquire an original geological texture image, screen and preprocess the original geological texture image to obtain a geological texture data set; Calculating the homogeneity, repeatability and directionality values ​​of each image sample in the geological texture dataset to obtain a sample index range, and performing sample screening on the geological texture dataset based on the sample index range to obtain a typical sample dataset; A geological texture synthesis network model based on a generative adversarial network is constructed, and the typical sample data set is input into the geological texture synthesis network model to obtain a synthetic geological texture image.

2. The method for synthesizing geological textures based on a GAN network according to claim 1, characterized in that: The calculating the homogeneity value of each image sample includes: Expanding the RGB channel values ​​of the image sample as the visual vocabulary of the image sample; Clustering the visual vocabulary using a K-means algorithm, merging words with similar meanings in the visual vocabulary, and generating a visual dictionary; According to the visual dictionary, the image samples are sampled in local areas of the same size, the underlying features of each local area are extracted and encoded, the labels of new data are determined according to the labels of surrounding neighboring data using a k-nearest neighbor algorithm, and the underlying features are quantized to the most similar visual vocabulary to construct a visual vocabulary vector of the local area; The Euclidean distance calculation is performed on the visual vocabulary vectors of any two local areas in the image sample, the calculation is repeated for a preset number of times, and the average value is used as the homogeneity value of the sample image.

3. The method for synthesizing geological textures based on a GAN network according to claim 1, characterized in that: The calculating of the repeatability value of each image sample includes: Perform correlation function calculation on the image sample and itself to obtain an autocorrelation matrix; Clipping the border of the autocorrelation matrix, and normalizing the clipped autocorrelation matrix to obtain a normalized autocorrelation matrix; A plurality of local areas in the image sample are randomly selected, a maximum value, a minimum value and a standard deviation of each local area are calculated, and the calculation results of the plurality of local areas are averaged to obtain a repeatability value of the image sample.

4. The method for synthesizing geological textures based on a GAN network according to claim 3, characterized in that: The calculation formula of the repeatability value of the image sample is: Among them, Max(Rt) and Min(Rt) represent the maximum and minimum values ​​of the random region Rt, respectively, σ(Min(Rt)) is the standard deviation of the minimum value, and T represents the number of randomly selected regions.

5. The method for synthesizing geological textures based on a GAN network according to claim 1, characterized in that: The calculating of the directionality value of each image sample comprises: Calculate the magnitude and angle of the gradient vector of each pixel of the image sample, and obtain the valid pixel point based on the preset gradient magnitude threshold; Dividing the angles of the effective pixel points into n types of directions, and obtaining a direction histogram of the sample image by statistics; The histogram of any category direction is expressed as: Among them, N θ (φ) represents the number of pixels with direction category φ in category n; The directivity value is calculated as: in, Indicates the direction Histogram of n p represents the number of peaks in the histogram, p represents the histogram peak, and w p Indicates the direction category range, Indicates the class of the peak in the histogram.

6. The method for synthesizing geological textures based on a GAN network according to claim 1, characterized in that: The construction of a geological texture synthesis network model based on a generative adversarial network includes: Add a multi-scale convolutional layer to the adversarial network to extract multi-scale features from the image and output a multi-scale feature map; A channel attention mechanism is added on the basis of multi-scale feature extraction, the multi-scale feature map is weighted and adjusted to obtain multi-scale weighted features, and the channel weights are redistributed according to the importance of each channel; Introducing shared source residual blocks, each residual block uses the same input source feature map and uses shared convolution weights to fuse features of different scales to learn multi-scale feature information; A content-aware upsampling module is introduced to combine the features of the input image to restore the detailed information of the geological texture.

7. The method for synthesizing geological textures based on a GAN network according to claim 6, characterized in that: The generator of the geological texture synthesis network model includes a multi-branch convolutional layer and a multi-scale attention module connected to the multi-branch convolutional layer; The multi-branch convolution layer is used to extract multi-scale features from the image. Each branch performs convolution calculation using grouped convolution to output a multi-scale feature map. The number of groups is dynamically adjusted according to the size of the convolution kernel. The multi-scale attention module is used to adjust the inter-channel information of the multi-scale feature map, obtain the channel attention weights of each scale, and integrate the multi-scale attention in a cascade manner to realize the interaction of attention information at multiple scales.

8. The method for synthesizing geological textures based on a GAN network according to claim 6, characterized in that: The content-aware upsampling module includes a sampling unit, a channel reduction unit, and a feature fusion unit connected in sequence; The sampling unit is used to perform 2-fold upsampling through CARAFE to obtain an upsampled feature map; The channel reduction unit uses Conv3×3-BN-LeakyReLU to reduce the channel; The feature fusion unit is used to perform an addition operation on the feature map after channel reduction and the original feature map to obtain a feature map that restores geological texture detail information.

9. The method for synthesizing geological textures based on a GAN network according to claim 1, characterized in that: The screening of the original geological texture image includes: Remove the images covered by interfering textures and keep the images with clear rock textures; Remove overexposed or shadowed images; Remove images where the angle between the shooting angle and the geological texture is greater than the inclination threshold.

10. The method for synthesizing geological textures based on a GAN network according to claim 1, characterized in that: The preprocessing includes image cropping and image culling; The image cropping process is used to crop the geological texture image into a uniform size; The image culling process is used to splice all geological texture images in the vertical and horizontal directions, and to cull texture samples that can be seamlessly spliced.

Images