A method and system for automatic partitioning of non-stationary training images based on deep features

Abstract

The present invention provides a method and system for automatically partitioning a non-stationary training image based on depth features, the method comprising: cutting and dividing a non-stationary training image into a plurality of training image sub-blocks; extracting the depth features of each training image sub-block based on a convolutional neural network model; performing cluster analysis on the depth features of all training image sub-blocks to obtain a clustering label of each training image sub-block; splicing each training image sub-block according to its position in the non-stationary training image based on the clustering label of each training image sub-block to obtain a partitioning result of the non-stationary training image; performing a stationary algorithm simulation based on the partitioning result to output a simulation implementation of the non-stationary training image. The present invention automatically partitions the non-stationary training image, and compared with the SIMPAT algorithm based on the non-stationary training image without partitioning, the multi-point geological statistical simulation results after the partitioning based on the present invention better reproduce the non-stationary features.

Description

Technical Field

[0001] The present invention relates to the field of reservoir geological modeling, and more specifically, to a method and system for automatically partitioning a non-stationary training image based on depth features. Background Art

[0002] Traditional geological statistics require the stationarity assumption, that is, any N-dimensional distribution function of the regionalized variable z(x) does not change with the displacement of the spatial point x. The stationarity assumption requires that the training image should satisfy stationarity, that is, the geometric configuration characteristics of the target body in the training image are basically unchanged throughout the region. The stationarity of the training image will directly affect the quality of the simulation results. The higher the stationarity, the closer the simulation is to the geological reality. Since changes in structure, provenance, etc. control the development of the sedimentary system in a non-stationary direction, most real geological conditions show strong non-stationarity, and typical non-stationary sedimentary phase types include the river delta.

[0003] Multi-point geostatistics solves the problem that traditional two-point statistical methods are difficult to characterize complex geological patterns. Multi-point statistical probabilities are established by scanning training images with data templates, and the obtained multi-point statistical probabilities are used to simulate the probability of the points to be estimated. Multi-point geostatistics does not use two-point statistical data like variograms, but borrows multi-point statistical data from training images. Therefore, multi-point geostatistics has advantages in reproducing complex geological patterns and overcomes the limitation that target-based methods are not easy to condition.

[0004] Multi-point geostatistics requires that the training images are stationary, that is, the sedimentary phase distribution density of the training images is uniform, and the structural characteristics are not strongly correlated with the spatial position. Stability exists in an ideal state, but there are many unstable geological structures in the actual natural world. For example, when a river flows out of the valley, the sediment breaks free from the lateral constraints and forms fan-shaped deposits. The sedimentary materials show non-stationary characteristics in space. At this time, classic multi-point geostatistics is difficult to characterize non-stationary reservoirs. Taking Simpat as an example, Figure 1 The Simpat method has a good reproduction effect on non-stationary features. If the non-stationarity in the training image is not considered, the simulation results will be difficult to reproduce the deposition morphology of the training image.

[0005] Existing non-stationary modeling methods are divided into three categories: the first category transforms non-stationary training images into stationary training images through scaling and rotation transformations, and then uses the obtained stationary training images for modeling. The second category associates spatial position and sedimentary features based on distance functions, considers the position information of the points to be estimated, and then performs non-stationary simulation. The third category is a partition-based method that divides non-stationary training images into relatively stationary sub-regions and performs simulations in the sub-regions. The above methods have disadvantages. The first category of modeling methods based on rotation and scaling of training images has difficulty in selecting rotation and scaling parameters, and the calculation is time-consuming, which increases the demand for computer memory. The second category of distance function methods based on non-stationary training images are sensitive to distance values ​​and difficult to determine the value. Among the third category of partition-based methods, the representative Gabor filter method has the disadvantage that the direction and frequency are difficult to define in advance.

[0006] Therefore, it is urgent to study a new method that can apply non-stationary training images for multi-point geostatistical modeling. Summary of the invention

[0007] The present invention aims at the technical problems existing in the prior art and provides a method and system for automatically partitioning non-stationary training images based on depth features, thereby providing a new method for applying non-stationary training images to perform multi-point geological statistical modeling.

[0008] According to a first aspect of the present invention, there is provided a method for automatically partitioning a non-stationary training image based on deep features, comprising:

[0009] Crop the non-stationary training image and divide it into a plurality of training image sub-blocks;

[0010] Extract deep features of each training image sub-block based on the convolutional neural network model;

[0011] Perform cluster analysis on the deep features of all training image sub-blocks to obtain the cluster label of each training image sub-block;

[0012] Based on the clustering label of each training image sub-block, each training image sub-block is spliced ​​according to the position in the non-stationary training image to obtain a partition result of the non-stationary training image;

[0013] Based on the partition result of the non-stationary training image, a stationary algorithm simulation is performed to output a simulated implementation of the non-stationary training image.

[0014] Based on the above technical solution, the present invention can also make the following improvements.

[0015] Optionally, the step of cropping the non-stationary training image into a plurality of training image sub-blocks includes:

[0016] Preprocessing the non-stationary training image, wherein the preprocessing includes at least image normalization, BM3D noise reduction and image enhancement;

[0017] Set the cropping size and step size of the sliding window;

[0018] Scanning the preprocessed non-stationary training image based on the sliding window, cutting the non-stationary training image into a plurality of training image sub-blocks.

[0019] Optionally, performing cluster analysis on the deep features of all training image sub-blocks to obtain a cluster label for each training image sub-block includes:

[0020] Determining the optimal number of clusters of the deep features of the training image sub-blocks based on the elbow rule, that is, the optimal number of partitions k of the non-stationary training image;

[0021] Based on the k-means clustering algorithm, the deep features of all training image sub-blocks are clustered in high-dimensional space to obtain the clustering label of each training image sub-block;

[0022] Based on the t-distributed random neighbor embedding algorithm, the deep features of each training image sub-block output after clustering analysis are reduced in dimension and visualized.

[0023] Optionally, determining the optimal number of clusters of the deep features of the training image sub-blocks based on the elbow rule, that is, the optimal number of partitions k of the non-stationary training image, includes:

[0024] The calculation formula of the elbow rule is:

[0025]

[0026] Where k is the number of clusters, that is, the number of partitions of the non-stationary training image, dist is the degree of distortion, x is the sample point, c i is the i-th cluster center;

[0027] The k value corresponding to the minimum SSE is taken as the optimal number of partitions for the non-stationary training images.

[0028] Optionally, the t-distribution random neighborhood embedding algorithm is used to perform dimension reduction visualization on the deep features of each training image sub-block output after cluster analysis, including:

[0029] The calculation formula of the t-distribution random neighborhood embedding algorithm is:

[0030]

[0031] Where x1,2,..., n is a set of N high-dimensional data, i.e., the deep features after intense analysis, N is the number of data samples, p ijFor data x i and x j The conditional probabilities of similarity between y1,2,..., m is a set of M low-dimensional data after dimensionality reduction, q ij for y i and j The conditional probability of being similar between

[0032] Visualize the deep features of each training image sub-block after dimensionality reduction.

[0033] Optionally, the extracting the deep features of each training image sub-block based on the convolutional neural network model includes:

[0034] Extracting the deep features of each training image sub-block based on multiple convolutional neural network models, and then obtaining the partition result of the non-stationary training image corresponding to each convolutional neural network model;

[0035] The step of simulating a stationary algorithm based on the partition result of the non-stationary training image and outputting a simulation implementation of the non-stationary training image comprises:

[0036] The classification accuracy of k-means clustering is evaluated for the deep features extracted by each convolutional neural network model, and the partition result corresponding to the convolutional neural network model with the highest classification accuracy is used as the final partition result of the non-stationary training image.

[0037] Optionally, the classification accuracy assessment of the deep features extracted by each convolutional neural network model by k-means clustering includes:

[0038] Based on the accuracy and precision of the classification results after k-means clustering, the classification accuracy of each convolutional neural network model is calculated. The accuracy refers to the proportion of correctly classified samples to the total number of samples, and the precision refers to the proportion of actually positive samples among all samples predicted to be positive examples.

[0039] Optionally, the step of splicing each training image sub-block according to a position in the non-stationary training image based on the clustering label of each training image sub-block to obtain a partition result of the non-stationary training image includes:

[0040] The cluster labels are assigned corresponding color labels, and based on the color label and the cluster label of each training image sub-block, each training image sub-block is spliced ​​according to its position in the non-stationary training image to obtain a partition result of the non-stationary training image.

[0041] According to a second aspect of the present invention, there is provided a system for automatically partitioning non-stationary training images based on deep features, comprising:

[0042] A cropping and partitioning module, used for cropping and partitioning a non-stationary training image into a plurality of training image sub-blocks;

[0043] An extraction module, used to extract deep features of each training image sub-block based on a convolutional neural network model;

[0044] A clustering analysis module is used to perform cluster analysis on the deep features of all training image sub-blocks to obtain a cluster label for each training image sub-block;

[0045] A splicing module, configured to splice each training image sub-block according to its position in the non-stationary training image based on the clustering label of each training image sub-block, so as to obtain a partitioning result of the non-stationary training image;

[0046] The simulation module is used to simulate the stationary algorithm based on the partition result of the non-stationary training image and output the simulation implementation of the non-stationary training image.

[0047] According to a third aspect of the present invention, there is provided an electronic device comprising a memory and a processor, wherein the processor is used to implement the steps of a method for automatically partitioning non-stationary training images based on deep features when executing a computer management program stored in the memory.

[0048] According to a fourth aspect of the present invention, there is provided a computer-readable storage medium having a computer management program stored thereon, which, when executed by a processor, implements the steps of a method for automatic partitioning of non-stationary training images based on deep features.

[0049] The present invention provides a method and system for automatically partitioning a non-stationary training image based on depth features. The entire non-stationary training image is cropped to obtain a number of training image sub-blocks. A pre-trained convolutional neural network model is used as a fixed feature extractor of the training image sub-blocks to extract the depth features of each training image sub-block. The extracted depth features are clustered to obtain a clustering label for each training image sub-block. All training image sub-blocks are spliced ​​based on the clustering labels to obtain a non-stationary training image partitioning result with a classification accuracy of 90.52%. The traditional stationary multi-point modeling algorithm is used for simulation within the partition to finally achieve non-stationary modeling. Therefore, compared with the SIMPAT algorithm based on the unpartitioned non-stationary training image, the multi-point geological statistical simulation results after partitioning based on the present invention better reproduce the non-stationary features. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A flow chart of a method for automatically partitioning a non-stationary training image based on deep features provided by the present invention;

[0051] Figure 2Schematic diagram of the simulation implementation of non-stationary training images and classic Simpat, where: Figure 2-1 is a schematic diagram of the original non-stationary training image. Figure 2-2 Schematic diagram of the simulation implementation reproduced by traditional methods;

[0052] Figure 3 Schematic diagram of cropping training images using sliding windows;

[0053] Figure 4 Schematic diagram of extracting the first 32-dimensional deep features of training image sub-blocks for the convolutional neural network DensNet121 model;

[0054] Figure 5 A diagram showing the relationship between the K value and the square distance error for calculating the depth feature of the training image sub-block based on the elbow rule;

[0055] Figure 6 Schematic diagram of clustering results after t-SNE dimension reduction based on K-means training image sub-block depth features;

[0056] Figure 7 A schematic diagram of the non-stationary training image partitioning results of the non-stationary training image automatic partitioning strategy based on deep features;

[0057] Figure 8 It is a flowchart of the automatic partitioning algorithm for non-stationary training images based on deep features;

[0058] Fig. 9 It is a schematic diagram of the multi-point geostatistical simulation results after the non-stationary training image automatic partitioning algorithm based on depth features is partitioned;

[0059] Fig.10 A schematic diagram of the structure of a non-stationary training image automatic partitioning system based on deep features provided by the present invention;

[0060] Fig.11 A schematic diagram of the hardware structure of a possible electronic device provided by the present invention;

[0061] Fig.12 A schematic diagram of the hardware structure of a possible computer-readable storage medium provided by the present invention. DETAILED DESCRIPTION

[0062] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0063] Absolute stationarity rarely exists in nature, and most geological phenomena exhibit non-stationary characteristics. Traditional geological statistics require the stationarity assumption to be met - assuming that geological phenomena are uniform in density and consistent in structural directionality. However, it is difficult to ensure that the training images are completely stationary in actual modeling. The present invention believes that the characteristics of non-stationarity are manifested as low-level visual features (such as structure and direction), and robust convolutional neural network models (such as VGG, ResNet, DenseNet) trained on large data sets can be used as deep feature extractors for non-stationary training images. After determining the number of sub-regions of the non-stationary training image, K-means clustering can classify the deep features of the training image with relative stationarity into one category to achieve the partitioning of the non-stationary training image.

[0064] The method proposed in the present invention is to transform the problem of training image partitioning into the problem of training image sub-block classification and mapping splicing, and automatically divide the entire training image into relatively stable training image sub-regions based on the deep features extracted from the training image sub-blocks, and all sub-regions are independent of each other.

[0065] Before describing the method proposed in the present invention, the relevant terms involved in the present invention are explained as follows:

[0066] 1. Training image (TI) is a two-dimensional or three-dimensional image that describes the geological structure to be modeled, used to represent the prior geological model concept.

[0067] 2. Simulation realization (R—Realization): The simulation model result is a digital model used to characterize the actual reservoir structure and its distribution pattern.

[0068] 3. Sliding Window: It is used to obtain the cropping range of the local image. Its size is determined by the sliding window height h and the sliding window width w. Step is the distance the sliding window center moves when the sliding window slides once.

[0069] 4. Training Image Sub-block: It is a local image, which is a local image of the training image obtained after sliding window cropping of the training image.

[0070] 5. SubRegion: A relatively stable subregion in a non-stationary training image.

[0071] 6. Convolutional Neural Networks: It is a type of feedforward neural network that includes convolutional layer operations and has a deep structure, that is, more than two hidden layers. It is one of the representative algorithms of deep learning.

[0072] 7. Deep Features: Features extracted based on deep network models.

[0073] 8. Elbow Methods: One of the methods for determining the K value in K-means clustering is the relationship function between the sum of squared errors (SSE-Sum of the Squared Errors) and k. SSE will decrease as K increases. When the decline rate tends to be slow, the K point, i.e. the elbow, is taken as the optimal K value.

[0074] 9. t-distributed Stochastic Neighbor Embedding (t-SNE): It is one of the best dimensionality reduction algorithms currently available. It converts the similarity between data points into conditional probabilities, uses Gaussian joint distribution to represent the similarity of data points in the original space, and the similarity of data points in the embedded space is represented by t-distribution.

[0075] 10. Accuracy: It refers to the ratio of correctly classified samples to the total number of samples. It is the simplest and most intuitive evaluation indicator in classification problems.

[0076] 11. Precision: It is the ratio of the number of samples that are actually positive examples among all the samples predicted to be positive examples.

[0077] 12. Micro Average: One of the ways to calculate precision is based on the weighted average of the number of samples, with each sample having the same weight.

[0078] 13. Macro Average: One of the ways to calculate precision is based on the weighted average of categories, with each category having the same weight.

[0079] Embodiment 1

[0080] A method for automatic partitioning of non-stationary training images based on deep features, see Figure 1 , the non-stationary training image automatic partitioning method mainly includes the following steps:

[0081] S1, cropping and dividing the non-stationary training image into multiple training image sub-blocks.

[0082] As an embodiment, a non-stationary training image is cropped and divided into a plurality of training image sub-blocks, including: preprocessing the non-stationary training image, wherein the preprocessing includes at least image normalization, BM3D noise reduction and image enhancement; setting a cropping size and a step size of a sliding window; and scanning the preprocessed non-stationary training image based on a sliding window, cropping the non-stationary training image into a plurality of training image sub-blocks.

[0083] It is understandable that the training images that multi-point geostatistics relies on require that the training images have stationarity, that is, the training images are uniform in density and consistent in structural direction. Stationarity exists in an ideal state. In the actual natural world, there are many unstable geological structures. At this time, classical multi-point geostatistics is difficult to characterize non-stationary reservoirs, such as Figure 2-1 and Figure 2-1 The training images shown contradict the simulation results.

[0084] The basis of traditional machine learning is to assume that the training data and the test data are in the same distribution, and the number of training data is sufficient to support model training. However, for multi-point geostatistics, there is not enough data set to support the training of a robust model. When the original data set is large enough, the model trained on this data set can learn enough image features to support the model to become a general model. Therefore, using a convolutional neural network pre-trained model on a large data set for feature extraction can solve the problem of insufficient training image samples. Non-stationary features appear as low-level visual features (such as structure and direction) to the naked eye, and deep network models in the field of image processing can be used to extract such features.

[0085] The training images need to be preprocessed before feature extraction, aiming to eliminate irrelevant information in the training images, highlight useful information, enhance the detectability of useful information and simplify the images to the maximum extent, thereby improving the reliability of feature extraction, image segmentation, matching and recognition. The training image preprocessing mainly includes image standardization, BM3D denoising, image enhancement, etc.

[0086] When deep feature extraction is performed on a training image, the features obtained are global features of the entire image. If you want to obtain local non-stationary features of a training image, you need to decompose it into training image sub-blocks. The embodiment of the present invention uses sliding window scanning to divide the non-stationary training image. Among them, the size of the sliding window is determined according to the specific phase type, the maximum value of the cropping size does not exceed one-third of the training image, and the minimum value of the cropping size depends on the phase type. The principle is that the phase structure can be distinguished by the naked eye after cropping. Take a 250*250 window and a step size of 25 to crop a fan delta training image as an example. The cropping process is as follows: Figure 3 shown.

[0087] S2, extracts the deep features of each training image sub-block based on the convolutional neural network model.

[0088] It can be understood that for each cropped training image sub-block, a convolutional neural network model is used to extract its deep features. The present invention builds a deep feature extractor based on the TensorFlow2.1 deep learning framework. The characterization of non-stationarity only requires low-level visual features (such as structure and direction), but does not require high-level visual features, such as the overall features extracted from the fully connected layer. Therefore, before loading the pre-trained model, the last fully connected layer needs to be deleted. Taking the use of the DensNet121 model to extract the deep features of the first training image sub-block as an example, the obtained partial feature map (32 dimensions) is shown in Figure 4 .

[0089] S3, performing cluster analysis on the deep features of all training image sub-blocks to obtain a cluster label for each training image sub-block.

[0090] As an embodiment, a cluster analysis is performed on the depth features of all training image sub-blocks to obtain a cluster label for each training image sub-block, including: determining the optimal number of clusters of the depth features of the training image sub-blocks based on the elbow rule, that is, the optimal number of partitions k of the non-stationary training image; performing a cluster analysis on the depth features of all training image sub-blocks in a high-dimensional space based on a k-means clustering algorithm to obtain a cluster label for each training image sub-block; and performing dimensionality reduction and visualization on the depth features of each training image sub-block output after the cluster analysis based on a t-distributed random neighborhood embedding algorithm.

[0091] It can be understood that after the non-stationary features of the training image sub-blocks are extracted, the deep features are obtained. The elbow rule is used to calculate the optimal clustering number k of the deep features, that is, the optimal partition number k of the non-stationary training image. The calculation result is as follows: Figure 5 As shown in Figure 1, performing K-means clustering on deep features can obtain non-stationary feature clustering results.

[0092] Specifically, the elbow rule is used to calculate the number of non-stationary training image partitions, that is, the optimal number of K-means clustering. The calculation formula of the elbow rule is:

[0093]

[0094] In the formula, k is the number of clusters, dist is the degree of distortion, x is the sample point, c is i is the i-th cluster center. According to different k values, the corresponding SEE value is calculated based on formula (1), and the k value corresponding to the minimum SSE value is taken as the optimal number of partitions for the non-stationary training image.

[0095] Determine the optimal partition number k value of the non-stationary training image, perform cluster analysis on the deep features of the extracted training image sub-blocks based on the k-means clustering algorithm, and obtain clustering results, including the clustering label of each training image sub-block (i.e., which category it belongs to) and the output clustering features. Among them, K-means clustering is used to cluster the deep features of the extracted training image sub-blocks, and the calculation formula of K-means clustering is:

[0096]

[0097] In the formula, n is the number of sample points, x i is the sample point, y i is the cluster center, dist(,Y) is the distance between each sample point data and the cluster center after clustering, and the clustering result corresponding to the minimum distance is taken as the best clustering result.

[0098] Among them, the feature output of K-means clustering is in a high-dimensional space, which needs to be visualized through dimensionality reduction. The t-distribution random neighbor embedding algorithm is used to reduce the dimensionality and visualize the K-means clustering results. The clustering visualization diagram is as follows: Figure 6 shown.

[0099] Specifically, the dimensionality reduction of deep features is performed based on the t-distributed random neighbor embedding algorithm. The calculation formula of the t-distributed random neighbor embedding algorithm is:

[0100]

[0101] Where x1,2,..., n is a set of N high-dimensional data, i.e., the deep features after cluster analysis, N is the number of data samples, p is ij For data x i and x j The conditional probabilities of similarity between y1,2,..., m is a set of M low-dimensional data after dimensionality reduction, q ij for y i and j The conditional probability of similarity between them is used to visualize the deep features after dimensionality reduction.

[0102] It should be noted that, as an embodiment, extracting the deep features of each training image sub-block based on the convolutional neural network model includes: extracting the deep features of each training image sub-block based on multiple convolutional neural network models, and then obtaining the partitioning result of the non-stationary training image corresponding to each convolutional neural network model; performing a stationary algorithm simulation based on the partitioning result of the non-stationary training image, and outputting a simulated implementation of the non-stationary training image, which includes: performing k-means clustering classification accuracy assessment on the deep features extracted by each convolutional neural network model, and taking the partitioning result corresponding to the convolutional neural network model with the highest classification accuracy as the final partitioning result of the non-stationary training image.

[0103] It is understandable that when extracting the deep features of each training image sub-block, a variety of different types of convolutional neural networks are used for extraction, and cluster analysis is performed based on the deep features extracted by different convolutional neural network models to obtain the partitioning results of the non-stationary training image. The classification accuracy of the deep features extracted by each convolutional neural network model is evaluated by k-means clustering, and the partitioning result corresponding to the convolutional neural network model with the highest classification accuracy is used as the final partitioning result of the non-stationary training image.

[0104] Among them, based on the accuracy and precision of the classification results, the classification accuracy of the classification results of each convolutional neural network model is calculated. In view of the non-stationary characteristics of the training image, it is currently difficult to determine which convolutional neural network model has the best effect. The present invention compares the accuracy of several typical convolutional neural network models, and the accuracy calculation formula is:

[0105]

[0106] Where n correct is the number of samples correctly classified, n total is the total number of samples. TP i To correctly predict the true classification i as the number of samples of classification i, F j P i is the number of samples that mistakenly predict the true category j as category i, then Table 1 lists the classification accuracy comparison of different convolutional neural network models (parts).

[0107] Table 1 Comparison of classification accuracy of different convolutional neural network models (partial)

[0108]

[0109] Densnet121 and Resnet50 have relatively higher accuracy and are closer, but the accuracy is mainly affected by the large number of categories. Therefore, it is necessary to calculate the accuracy of the two on this basis to more objectively measure the classification quality of the two networks. The accuracy is the proportion of samples that are actually positive examples among all samples predicted to be positive examples. There are two calculation methods: macro and micro. Macro is a weighted average based on categories, and each category has the same weight; micro is a weighted average based on the number of samples, and each sample has the same weight. The calculation formulas are:

[0110]

[0111]

[0112] Where P i is the prediction accuracy of category i, Table 2 lists the accuracy and precision calculation results of Densnet121 and Resnet50.

[0113] Table 2 Classification accuracy and precision of training images Densnet121 and Resnet50

[0114]

[0115]

[0116] It can be seen from the above table that the accuracy and precision of Densnet121 are better than Resnet50.

[0117] After the optimal convolutional neural network model is determined, the training image will be partitioned, and the file names of all training image sub-blocks will be bound to the cluster labels after K-means clustering and stored in the cluster label dictionary.

[0118] S4, based on the clustering label of each training image sub-block, splicing each training image sub-block according to its position in the non-stationary training image to obtain a partition result of the non-stationary training image.

[0119] As an embodiment, based on the clustering label of each training image sub-block, each training image sub-block is spliced ​​according to its position in the non-stationary training image to obtain a partitioning result of the non-stationary training image, including: assigning corresponding color labels to the clustering labels, and based on the color label and clustering label of each training image sub-block, each training image sub-block is spliced ​​according to its position in the non-stationary training image to obtain a partitioning result of the non-stationary training image.

[0120] It can be understood that for each training image sub-block, a color label is defined according to its clustering label, and a color map with the optimal number of partitions is created; a color is selected as a sub-area label and assigned to the training image sub-block according to the clustering label, and the same color is assigned to the training image sub-blocks with the same clustering label, that is, different colors are used to distinguish different categories of training image sub-blocks; the marked training image sub-blocks are spliced ​​according to position to obtain the partition result of the non-stationary training image. In the process of splicing, a picture with the same size as the original training image is defined, and a row-column nested loop is performed to perform mapping splicing to obtain the result of the non-stationary partition, such as Figure 7 The whole algorithm flow chart is as shown in Figure 8 shown.

[0121] S5, based on the partition result of the non-stationary training image, perform a stationary algorithm simulation, and output a simulated implementation of the non-stationary training image.

[0122] It can be understood that after the partition result of the non-stationary training image is obtained, the classical stationary algorithm simulation is performed in the relatively stable sub-region, and the output simulation is realized. The non-stationary training image ( Figure 2-1 ) partition results were used for multi-point geostatistical simulation, simulating two random realizations ( Fig. 9 ), simulation results and implementation of traditional multi-point geostatistical Simpat simulation based on unpartitioned non-stationary training images ( Figure 2-2 ), it well reproduces the river channel sedimentary characteristics of the fan delta.

[0123] Embodiment 2

[0124] An automatic partitioning system for non-stationary training images based on deep features, see Fig.10 The non-stationary training image automatic partitioning system includes a cropping module 101, an extraction module 102, a clustering analysis module 103, a splicing module 104 and a simulation module 105, wherein:

[0125] The cropping module 101 is used to crop the non-stationary training image and divide it into multiple training image sub-blocks; the extraction module 102 is used to extract the depth features of each training image sub-block based on the convolutional neural network model; the clustering analysis module 103 is used to perform clustering analysis on the depth features of all training image sub-blocks to obtain the clustering label of each training image sub-block; the splicing module 104 is used to splice each training image sub-block according to its position in the non-stationary training image based on the clustering label of each training image sub-block to obtain the partition result of the non-stationary training image; the simulation module 105 is used to simulate the stationary algorithm based on the partition result of the non-stationary training image, and output the simulated implementation of the non-stationary training image.

[0126] It can be understood that the automatic partitioning system for non-stationary training images based on depth features provided by the present invention corresponds to the automatic partitioning method for non-stationary training images based on depth features provided in the aforementioned embodiments. The relevant technical features of the automatic partitioning system for non-stationary training images based on depth features can refer to the relevant technical features of the automatic partitioning method for non-stationary training images based on depth features, and will not be repeated here.

[0127] Embodiment 3

[0128] See also Fig.11 , Fig.11 Schematic diagram of an electronic device provided by an embodiment of the present invention. Fig.11 As shown, an embodiment of the present invention provides an electronic device 1100, including a memory 1110, a processor 1120, and a computer program 1111 stored in the memory 1110 and executable on the processor 1120. When the processor 1120 executes the computer program 1111, the steps of the method for automatic partitioning of non-stationary training images based on deep features of embodiment 1 are implemented.

[0129] Embodiment 4

[0130] See also Fig.12 , Fig.12 Schematic diagram of an embodiment of a computer-readable storage medium provided by the present invention. Fig.12 As shown, this embodiment provides a computer-readable storage medium 1200 on which a computer program 1211 is stored. When the computer program 1211 is executed by a processor, the steps of the method for automatic partitioning of non-stationary training images based on deep features in embodiment 1 are implemented.

[0131] An embodiment of the present invention provides a method and system for automatically partitioning a non-stationary training image based on depth features. The entire non-stationary training image is cropped to obtain a number of training image sub-blocks, and a pre-trained convolutional neural network model is used as a fixed feature extractor of the training image sub-block to extract the depth features of each training image sub-block. The extracted depth features are clustered to obtain a clustering label for each training image sub-block, and all training image sub-blocks are spliced ​​based on the clustering label to obtain a non-stationary training image partitioning result with a classification accuracy of 90.52%. The traditional stationary multi-point modeling algorithm is used for simulation within the partition to finally achieve non-stationary modeling. Therefore, compared with the SIMPAT algorithm based on the unpartitioned non-stationary training image, the multi-point geological statistical simulation results after partitioning based on the present invention better reproduce the non-stationary features.

[0132] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and for parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0133] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0134] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0135] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0136] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0137] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0138] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for automatic partitioning of non-stationary training images based on deep features, characterized in that: include: Crop the non-stationary training image and divide it into a plurality of training image sub-blocks; Extract deep features of each training image sub-block based on the convolutional neural network model; Perform cluster analysis on the deep features of all training image sub-blocks to obtain the cluster label of each training image sub-block; Based on the clustering label of each training image sub-block, each training image sub-block is spliced ​​according to the position in the non-stationary training image to obtain a partition result of the non-stationary training image; Based on the partition result of the non-stationary training image, a stationary algorithm simulation is performed to output a simulated implementation of the non-stationary training image; The clustering analysis is performed on the deep features of all training image sub-blocks to obtain a cluster label for each training image sub-block, including: Determining the optimal number of clusters of the deep features of the training image sub-blocks based on the elbow rule, that is, the optimal number of partitions k of the non-stationary training image; Based on the k-means clustering algorithm, the deep features of all training image sub-blocks are clustered in high-dimensional space to obtain the clustering label of each training image sub-block; Based on the t-distributed random neighborhood embedding algorithm, the deep features of each training image sub-block output after clustering analysis are reduced in dimension and visualized; The method of determining the optimal clustering number of the deep features of the training image sub-blocks based on the elbow rule, that is, the optimal partition number k of the non-stationary training image, includes: The calculation formula of the elbow rule is: ； In the formula, is the number of clusters, i.e. the number of partitions of the non-stationary training images, is the degree of distortion, is the sample point, is the i-th cluster center; Minimize The corresponding k value is taken as the optimal number of partitions for the non-stationary training images; The t-distribution random neighborhood embedding algorithm is used to reduce the dimension of the deep features of each training image sub-block output after cluster analysis, and visualize the depth features, including: The calculation formula of the t-distribution random neighborhood embedding algorithm is: ； In the formula, is a set of N high-dimensional data, i.e., the deep features after intense analysis, where N is the number of data samples. For data and The conditional probability of being similar between is a set of M low-dimensional data after dimensionality reduction, for and The conditional probability of being similar between Visualize the deep features of each training image sub-block after dimensionality reduction.

2. The method for automatically partitioning non-stationary training images according to claim 1, characterized in that: The step of cropping the non-stationary training image into a plurality of training image sub-blocks includes: Preprocessing the non-stationary training image, wherein the preprocessing includes at least image normalization, BM3D noise reduction and image enhancement; Set the cropping size and step size of the sliding window; Scanning the preprocessed non-stationary training image based on the sliding window, cutting the non-stationary training image into a plurality of training image sub-blocks.

3. The method for automatic partitioning of non-stationary training images according to claim 1, characterized in that: The method of extracting the deep features of each training image sub-block based on the convolutional neural network model includes: Extracting the deep features of each training image sub-block based on multiple convolutional neural network models, and then obtaining the partition result of the non-stationary training image corresponding to each convolutional neural network model; The step of simulating a stationary algorithm based on the partition result of the non-stationary training image and outputting a simulation implementation of the non-stationary training image comprises: The classification accuracy of k-means clustering is evaluated for the deep features extracted by each convolutional neural network model, and the partition result corresponding to the convolutional neural network model with the highest classification accuracy is used as the final partition result of the non-stationary training image.

4. The method for automatically partitioning non-stationary training images according to claim 3, characterized in that: The classification accuracy assessment of the k-means clustering of the deep features extracted by each convolutional neural network model includes: Based on the accuracy and precision of the classification results after k-means clustering, the classification accuracy of each convolutional neural network model is calculated. The accuracy refers to the proportion of correctly classified samples to the total number of samples, and the precision refers to the proportion of actually positive samples among all samples predicted to be positive examples.

5. The method for automatically partitioning non-stationary training images according to claim 1, characterized in that: The method of splicing each training image sub-block according to its position in the non-stationary training image based on the clustering label of each training image sub-block to obtain a partition result of the non-stationary training image includes: The cluster labels are assigned corresponding color labels, and based on the color label and the cluster label of each training image sub-block, each training image sub-block is spliced ​​according to its position in the non-stationary training image to obtain a partition result of the non-stationary training image.

6. A non-stationary training image automatic partitioning system based on deep features, characterized in that: include: A cropping and partitioning module, used for cropping and partitioning a non-stationary training image into a plurality of training image sub-blocks; An extraction module, used to extract deep features of each training image sub-block based on a convolutional neural network model; A clustering analysis module is used to perform cluster analysis on the deep features of all training image sub-blocks to obtain a cluster label for each training image sub-block; A splicing module, configured to splice each training image sub-block according to its position in the non-stationary training image based on the clustering label of each training image sub-block, so as to obtain a partitioning result of the non-stationary training image; A simulation module, used for performing a stationary algorithm simulation based on the partition result of the non-stationary training image, and outputting a simulation implementation of the non-stationary training image; The clustering analysis is performed on the deep features of all training image sub-blocks to obtain a cluster label for each training image sub-block, including: Determining the optimal number of clusters of the deep features of the training image sub-blocks based on the elbow rule, that is, the optimal number of partitions k of the non-stationary training image; Based on the k-means clustering algorithm, the deep features of all training image sub-blocks are clustered in high-dimensional space to obtain the clustering label of each training image sub-block; Based on the t-distributed random neighborhood embedding algorithm, the deep features of each training image sub-block output after clustering analysis are reduced in dimension and visualized; The method of determining the optimal clustering number of the deep features of the training image sub-blocks based on the elbow rule, that is, the optimal partition number k of the non-stationary training image, includes: The calculation formula of the elbow rule is: ； In the formula, is the number of clusters, i.e. the number of partitions of the non-stationary training images, is the degree of distortion, is the sample point, is the i-th cluster center; Minimize The corresponding k value is taken as the optimal number of partitions for the non-stationary training images; The t-distribution random neighborhood embedding algorithm is used to reduce the dimension of the deep features of each training image sub-block output after cluster analysis, and visualize the depth features, including: The calculation formula of the t-distribution random neighborhood embedding algorithm is: ； In the formula, is a set of N high-dimensional data, i.e., the deep features after intense analysis, where N is the number of data samples. For data and The conditional probability of being similar between is a set of M low-dimensional data after dimensionality reduction, for and The conditional probability of being similar between Visualize the deep features of each training image sub-block after dimensionality reduction.

Images