Typical interlayer oxidation zone zoning identification method and device for sandstone type uranium deposit

Abstract

The invention discloses a sandstone type uranium mine typical interlayer oxidation zone zoning identification method based on a convolutional neural network, and the method comprises the following steps: 1, obtaining a rock core cross section image, determining a zoning label and a granularity label, and making a sample data set; 2, preprocessing by adopting a digital image technology, and dividing into a training set and a test set; 3, constructing and training a model: selecting a visual classification neural network model, and starting deep learning training of the model by using the preprocessed training set; 4, evaluating a model training result, selecting an optimal prediction model, and verifying the optimal prediction model by adopting a cross validation method; and step 5, applying model prediction. According to the method, the acquired image can be identified through the convolutional neural network to judge the category and granularity in the interlayer oxidation zone to which the image belongs, so that the technical problem that the sandstone type uranium ore zone is not accurately identified manually in the prior art is effectively solved, and the identification efficiency and accuracy are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision and image recognition, and in particular relates to a method and a device for identifying typical interlayer oxidation zones of sandstone-type uranium deposits based on a convolutional neural network. Background Art

[0002] Sandstone-type uranium deposits have an important position and broad application prospects in mineral exploration and development in the world and China due to their large resource volume and convenient mining. However, since sandstone-type uranium deposits are mostly hidden underground, the extraction of exploration indicators for sandstone-type uranium deposits has become an important research content for the exploration of such deposits. At present, the methods that can relatively accurately analyze sandstone-type uranium deposits mainly include traditional optical microscope observation, X-ray diffraction experiment (XRD), TESCAN automatic mineral analysis system (TIMA) and other methods. However, traditional optical microscopes require observers to have sufficient observation experience and the accuracy and speed of naked eye observation have great individual differences; although XRD and TIMA can obtain detailed mineral composition and content, particle morphology and size, etc., their expensive testing costs and the scarce number of testing instruments require more preliminary preparation work to use this method to ensure that the data obtained from the test have sufficient reference significance.

[0003] In the study of interlayer oxidation zones of sandstone-type uranium deposits, various theories have been proposed, as well as the study of the mineralization process of sandstone-type uranium deposits and different zoning of cores. The mineralization process is affected by groundwater conditions, and the aquifer can be divided into oxidation zone, redox transition zone (metallogenic zone) and reduction zone.

[0004] With the continuous development of computer technology and artificial intelligence, computer vision technology, including digital image processing and deep neural network technologies, is increasingly used in lithology identification. It has many significant advantages: first, it can significantly improve exploration efficiency and accuracy, and use a large number of core images to train convolutional neural network models, forming a lower-cost and more efficient research method, effectively avoiding resource waste; second, it helps to make up for the shortcomings in existing research. The current research on deep learning models in the geological field focuses on making the model adaptable to the identification of more types of rocks and minerals. However, the types of rocks and minerals in nature are extremely complex, with a wide variety of shapes and properties. It is difficult to rely solely on image data to carry enough information density for machine learning and classification. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for identifying typical interlayer oxidation zones of sandstone-type uranium deposits, which can identify the acquired images through a convolutional neural network to determine the category and particle size of the interlayer oxidation zone to which it belongs, and can improve the recognition efficiency and accuracy, and provide an identification device for deploying the algorithm.

[0006] The object of the present invention is achieved through the following technical scheme: A method for identifying typical interlayer oxidation zones of sandstone-type uranium deposits, comprising the following steps:

[0007] Step 1: Obtain a core cross-section image of a typical interlayer oxidation zone sandstone-type uranium deposit, compare the core cross-section image with the spatial position information of the sample, and determine the stratigraphic layer information to which the sample belongs, and then determine the zoning label of the sample; determine the sandstone grain size by observing the grain size and distribution characteristics of the sandstone core, and produce a sample data set containing zoning labels and grain size labels;

[0008] Step 2: Preprocess the images in the sample data set using digital image technology, and divide the sample data set into a training set and a test set;

[0009] Step 3: Build and train the model: Select a variety of different visual classification neural network models and use the preprocessed training set to start the deep learning training of the model;

[0010] Step 4: Evaluate and verify the model: Evaluate the model training results and select the optimal prediction model;

[0011] Step 5: Apply model prediction: Use the optimal prediction model to predict the zoning type of the target sandstone-type uranium ore sample image and visualize the prediction results.

[0012] Another object of the present invention is to provide a typical interlayer oxidation zone identification device for sandstone-type uranium ore. The above method realizes the typical interlayer oxidation zone identification of sandstone-type uranium ore, and includes the following modules:

[0013] Acquisition module: responsible for collecting images and recording banding and granularity labels to form a data set;

[0014] Preprocessing module: Use digital image technology and data preprocessing methods to process images in the data set;

[0015] Training module: select different models in the image classification toolbox, write or modify configuration files, and call pre-trained models for deep learning training;

[0016] Verification module: evaluates model training results, selects the best prediction model and performs cross-validation;

[0017] Prediction module: Use the optimal prediction model to predict lithology and visualize the prediction results.

[0018] The beneficial effect of the present invention is that the present invention can recognize the acquired image through a convolutional neural network to determine the category and particle size of the interlayer oxidation zoning to which it belongs, effectively solving the technical problem of inaccurate manual identification of sandstone-type uranium ore zoning in the prior art, and can improve the recognition efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A schematic flow chart of a typical interlayer oxidation zone identification method for sandstone-type uranium deposits based on a convolutional neural network provided by the present invention;

[0020] Figure 2 This is a core cross-section image and classification label of a typical interlayer oxidation zone type uranium deposit in Yili, Xinjiang;

[0021] Figure 3 A visualization diagram of the attention mechanism for the sandstone zoning type classification results provided in this embodiment;

[0022] Figure 4 Evaluation chart of prediction results of different visual classification neural network models provided for this example;

[0023] Figure 5 The prediction results and predicted zoning probability diagram of different zoning samples of sandstone-type uranium deposits using the optimal model provided in this embodiment. DETAILED DESCRIPTION

[0024] In order to address the difficulties in integrating big data artificial intelligence with geology and using convolutional neural network models to process sandstone-type uranium deposit core images, we can construct a data set by collecting core samples from specific deposits, deeply analyze the correlation between mineral composition differences and image data, build and optimize the neural network model, achieve intelligent recognition and prediction functions, and test the reliability of the model.

[0025] The technical solution of the present invention is further described below in conjunction with the accompanying drawings. Figure 1 As shown, a typical interlayer oxidation zone identification method for sandstone-type uranium deposits based on a convolutional neural network of the present invention comprises the following steps:

[0026] Step 1, obtain a core cross-sectional image of a typical interlayer oxidation zone sandstone-type uranium deposit, wherein the core cross-sectional image taken must ensure accurate focus, clear shooting, appropriate brightness, and the core is not contaminated by impurities such as mud. In this embodiment, a core cross-sectional photo is obtained by taking photos in a core library of a typical interlayer oxidation zone sandstone-type uranium deposit (Mengqigur sandstone-type uranium deposit, Honghaigou sandstone-type uranium deposit) in Yili, Xinjiang, using equipment (including but not limited to digital cameras, mobile phones or other professional equipment).

[0027] Then, the following operations are performed on the captured core cross-section image:

[0028] If there is a reference natural gamma curve in the borehole to which the core cross-section image belongs, the sample image can be judged whether it belongs to mineral-bearing sandstone based on the abnormal condition of the curve. Specifically, when the peak position of the natural gamma curve reaches more than 2 times the background average gamma value (such as the average background gamma value is 20 (nC / (kg*h)), and the logging natural gamma value is above 40 (nC / (kg*h)), this peak position is the uranium anomaly point, which means it is a transition zone.

[0029] If there is no natural gamma curve of the borehole, the specific location information of different zoning layers in the typical interlayer oxidation zone can be determined by integrating the information of multiple boreholes into a cross-sectional geological map with information on oxidation zones, transition zones (ore-bearing zones), and reduction zones.

[0030] Finally, the various types of information obtained above (including information related to the natural gamma curve, information on different zoning layers, etc.) are compared with the spatial location information of the sample to determine the stratigraphic layer information to which the sample belongs, and then determine the zoning label of the sample. These testing methods must ensure accuracy and scientificity to accurately mark the attributes of the core cross-section photos. Figure 2 Image data and classification labels constructed according to an embodiment of the present invention.

[0031] The method for determining the grain size label is to observe the hand specimens with the naked eye and the rock slices under a microscope to observe the particle size, distribution and other characteristics of the sandstone core to determine the sandstone grain size. Finally, a sample data set containing zoning labels and grain size labels is produced. The obtained core cross-section photos must cover image content of different zoning and different grain sizes, specifically including relatively coarse-grained coarse sandstone to gravel-bearing coarse sandstone in the ore-bearing section, reduction zone, and oxidation zone sandstone, as well as relatively fine-grained fine sandstone to siltstone. In view of the fact that coarse-grained coarse sandstone to gravel-bearing coarse sandstone is difficult to see in the samples of the reduction zone, the samples in this embodiment include the following five categories: ore-bearing coarse sandstone, ore-bearing fine sandstone, reduction zone fine sandstone, oxidation zone fine sandstone, oxidation zone coarse sandstone.

[0032] Step 2: Use digital image technology to preprocess the images in the sample data set, and divide the sample data set into a training set and a test set; the specific implementation method is:

[0033] Step 21: Perform format conversion: Perform format conversion (such as conversion between JPG, PNG, YUV, and between RGB, LAB, and HSV) to meet different processing requirements. For example, some images taken by some devices are not in JPG or PNG format that can be used for deep learning model training, but in HELF and other formats. In this step, the image format needs to be uniformly converted into JPG or PNG format that can be suitable for model training; in some image processing algorithms, it may be necessary to convert the image to a specific color space (such as LAB space) to better analyze the color characteristics of the image.

[0034] Step 22: Core image segmentation: For each image in the sample data set, the pre-trained FasterRCNN image segmentation model is used, supplemented by manual evaluation of the segmentation effect, to extract the core part from the image; the FasterRCNN model has efficient target detection and segmentation capabilities. It learns the features in the image through convolutional neural networks and can accurately identify the position of the core in the image. Combined with manual evaluation, it can further improve the accuracy and reliability of segmentation.

[0035] Step 23, Image cutting: Cut the extracted image containing only the core part into an image with a resolution of 256*256; this specific resolution selection helps to improve the computational efficiency and image quality stability in subsequent processing and analysis. The unified image size and format facilitates the model to process image data and reduce the interference caused by differences in image size and format.

[0036] Step 24: Use denoising techniques (such as fast mean method, histogram equalization, etc.) to remove noise interference in the image; noise may come from factors such as shooting equipment and environment, which will affect the clarity and feature extraction of the image. The fast mean method smoothes the image and reduces noise by calculating the average value of the local neighborhood of the image; histogram equalization enhances image contrast by stretching the image pixel intensity distribution, and can also reduce noise to a certain extent.

[0037] Step 25: Use repair technology (such as Navier-Stokes crack repair technology) to repair defects such as cracks in the image and restore the integrity of the image. If there are cracks and other damaged parts in the image, it may affect the observation and analysis of the core characteristics. The repair technology can make the image more complete, which is conducive to subsequent processing.

[0038] Step 26: Perform image enhancement: Perform image enhancement operations such as image flipping, rotation, and brightness adjustment to expand the diversity and richness of the data set. These operations can change the image's viewing angle, brightness, and other attributes, increase the diversity of the data, enable the model to learn more different image features, and improve the model's generalization ability.

[0039] Step 27, data set division: divide the sample data set after image enhancement into a training set and a test set. In this embodiment, 80% of the images in the image set processed by digital image technology are randomly divided as the Train training set, and the remaining 20% ​​are used as the Val test set. This division method can ensure that there is sufficient training data when training the model, and the performance of the model can be objectively evaluated through the test set. The training set is used to train the model to allow the model to learn the relationship between image features and banding and granularity labels; the test set is used to verify the accuracy and generalization ability of the model. It does not participate in the training during the model training process and is only used to evaluate the performance of the model on unseen data.

[0040] Step 3: Build and train the model: Select a variety of different visual classification neural network models, write or modify the configuration file, call the pre-trained model of the selected visual classification neural network model, and use the pre-processed training set to start the deep learning training of the model;

[0041] Model selection: Choose different visual classification neural network models, including ResNet, ResNeXt, MobileNet, DenseNet, EfficientNet, MViT, VGG, and GoogleNet (inception). These models have a wide range of applications in the field of computer vision. Train these models separately and select the best model from them in the subsequent steps.

[0042] Configuration file writing and modification: Write or modify the configuration file, which includes the specified dataset path, model structure, learning rate strategy, optimizer, etc.

[0043] Dataset path: Explicitly tell the model where the training data is located so that the model can correctly read the data for training.

[0044] Model structure: According to the selected neural network model, determine its specific structural parameters, such as the number of layers, the number of neurons in each layer, the size of the convolution kernel, etc. These parameters determine the complexity and learning ability of the model.

[0045] Learning rate strategy: The learning rate is an important parameter in the model training process, which controls the step size of the model weight update. A suitable learning rate strategy can make the model converge to the optimal solution faster. Common learning rate strategies include fixed learning rate, learning rate decay, etc. For example, a larger learning rate can be used in the early stage of training to quickly adjust the model weights. As the training progresses, the learning rate can be gradually reduced to make the model converge more stably.

[0046] Optimizer: Select a suitable optimizer to update the model weights. Common optimizers include stochastic gradient descent (SGD), Adagrad, Adadelta, Adam, etc. Different optimizers have different characteristics and applicable scenarios. For example, the Adam optimizer combines the advantages of Adagrad and RMSProp and can achieve good performance in most cases. It can adaptively adjust the learning rate and update the weights according to the gradient information of each parameter. Therefore, the optimizer selected in this example is the Adam optimizer.

[0047] Model training: call different visual classification neural network pre-trained models, and start deep learning training on the preprocessed data set. The pre-trained model is a model that has been pre-trained on a large-scale image data set, and it has learned some common image features. By using a pre-trained model, the training speed of the model in the present invention can be accelerated, and the performance of the model can be improved to a certain extent. During the training process, the model continuously adjusts its own weights according to the input image data and the corresponding banding and granularity labels, so that the error between the prediction results output by the model and the true label is minimized. The model calculates the gradient through the back-propagation algorithm, and updates the weights according to the selected optimizer and learning rate strategy. After multiple iterative training, the accuracy and generalization ability of the model are gradually improved.

[0048] The learning rate is a parameter that controls the weight update step size during model training. Its calculation formula is: where θ new and θ represent the new weight and the current weight respectively, η is the learning rate, is the gradient of the loss function with respect to the weight θ; a suitable learning rate can make the model converge to the optimal solution faster.

[0049] Model Interpretability Feedback: In the present invention, the class activation mapping (CAM) technique is used to perform interpretability analysis on the trained visual classification neural network. By generating heat maps for the input image and its corresponding prediction results, CAM provides an intuitive method to reveal the areas that the model focuses on during the classification decision process. Specifically, CAM uses the feature map output by the convolutional layer and its corresponding weight information to calculate the significant feature areas that affect the classification results, thereby generating a visual image with heat information. The implementation of this process includes the following steps: first, the image to be analyzed is input into the trained model and the class probability of its output is obtained; then, the feature map of the last convolutional layer in the model and the corresponding classification weight are extracted. By weighting the feature map and applying the activation function, the generated heat map reflects the model's assessment of the importance of different image regions. This heat map can effectively indicate the decision basis of the model, allowing users to identify the key visual features that the model relies on under specific predictions. In the feedback analysis, CAM not only enhances the interpretability of the model, but also provides a basis for model optimization. By comparing the overlap between the heat map and the real annotation, researchers can identify the potential bias and misclassification reasons of the model, thereby further optimizing the training process and data set, and improving the overall performance and reliability of the model. This feedback mechanism has important guiding significance in model development and application. Figure 3 As shown in the figure, the leftmost part is the input core image, which is the original data for the model to classify; the middle part is the CAM heat map, which is generated based on the attention mechanism. The redder the area, the more focused the neural network is on the area. The color depth intuitively reflects the difference in the degree of attention paid by the model to different areas. This method can effectively reveal the key areas that the model focuses on when making classification decisions; the far right is the superposition of the input core image and the heat map. This superposition effect helps to more clearly observe the specific parts of the image that the model pays attention to when classifying sandstone zoning types, and then deeply understand the decision-making basis of the model.

[0050] Step 4: Evaluate and verify the model: Evaluate the model training results and select the optimal prediction model; the specific implementation method is:

[0051] Step 41, determine the evaluation index: evaluate by comparing the prediction results of different visual classification neural network models, the evaluation index includes TOP1 prediction rate (accuracy) and loss function (loss);

[0052] The TOP1 prediction rate is calculated as the number of samples S that match the predicted label and the actual label. top1 Divide by the total sample size S t , that is, S top1 / S t, which reflects the correct proportion of the most likely category predicted by the model. This evaluation index intuitively quantifies the accurate proportion of the model's determination of the zoning and grain size categories of sandstone-type uranium deposit core images.

[0053] The loss function is a common loss function such as the cross entropy loss function, and its calculation formula is: Where m represents the number of categories, y ic is an indicator variable (valued as 0 or 1, indicating whether the observed sample i belongs to category c, if yes, then it is 1, otherwise it is 0), is the predicted probability that the observed sample i belongs to category c; the smaller the value of the loss function, the closer the model's prediction result is to the true label. In actual mineral deposits, sandstone samples of different zones (such as oxidation zone, reduction zone, mineral-bearing zone) and grain sizes (coarse sandstone, fine sandstone, etc.) contain unique geological characteristics and mineral combination information, which correspond to different category labels. When the model predicts the category to which a sandstone core image belongs, the predicted probability distribution If it deviates from the true category (by y ic Indication), the loss function value increases.

[0054] Step 42, model selection: According to the above evaluation indicators, select the prediction model that meets the conditions as the candidate model; input the test set data into different visual classification neural network models, calculate the TOP1 prediction rate and loss function of each model respectively, and select the model whose TOP1 prediction rate is stably greater than 0.8 and the loss function is stably less than 0.3 as the number of training rounds gradually increases for cross-validation; Figure 4 Figure 2 is a graph of the loss function and accuracy of some models.

[0055] Step 43, cross validation: Use K-fold cross validation method to verify the generalization ability of the model and select the optimal model. The cross validation method includes the following steps:

[0056] Dataset segmentation: The entire data set is divided into k mutually exclusive subsets of similar size; in the example of the present invention, k=5, that is, the sandstone-type uranium ore core image data set is evenly divided into 5 parts, the data set contains 1000 core cross-section images, each subset has 200 core images, and the division adopts the random division method to ensure that the data in each subset is random and representative. In the division process, it is necessary to keep the proportion of images of different zoning grain sizes (ore-bearing coarse sandstone, ore-bearing fine sandstone, reduction zone fine sandstone, oxidation zone fine sandstone, oxidation zone coarse sandstone) similar in each subset to avoid some subsets concentrating too many images of a certain category, thereby affecting the evaluation results.

[0057] Training and validation: Perform 5 training and validation cycles. Each time, select 1 subset as the validation set, and combine the remaining 4 subsets as the training set. For example, the first time, use the 1st subset as the validation set, and use the 2nd to 5th subsets as the training sets for training. The second time, use the 2nd subset as the validation set, and use the 1st, 3rd to 5th subsets as the training sets for training, and so on.

[0058] Model evaluation: Evaluate the overall performance and stability of the model through prediction results; calculate the evaluation indicators (TOP1 prediction rate, loss function) on each validation set, and combine these results to judge the performance of the model on different data subsets, thereby evaluating the generalization ability of the model.

[0059] Among them, the specific method for evaluating the generalization ability of the model is as follows: (1) Stability evaluation: observe the fluctuation of the TOP1 prediction rate in k validations. If the prediction rate fluctuates slightly, for example, it fluctuates between 0.85-0.95, it means that the performance of the model on different data subsets is relatively stable; on the contrary, if the fluctuation is large, such as the prediction rate is 0.9 in one validation and 0.6 in another, it means that the stability of the model is poor, and it may be more sensitive to certain characteristics of the data, and the generalization ability is at risk. (2) Average performance evaluation: calculate the average value of the loss function in k validations. The smaller the average loss function value, the smaller the difference between the prediction results of the model in different subsets and the true label, and the more reliable the model prediction.

[0060] Among them, for the application scenario of typical interlayer oxidation zone identification of sandstone-type uranium deposits, stability is the primary consideration. Even if the average performance of the model is slightly weak, as long as it can stably give reliable results, it can avoid too many misjudgments due to large fluctuations in the results. Therefore, when the average performance indicators of multiple models are within a better range and the average value of the loss function is less than 0.3, it is considered that the average performance of the model is already at a sufficiently high level. Among these models with relatively good average performance, the one with the best stability evaluation effect is selected. For example, the average loss function values ​​of models A, B, and C are 0.25, 0.2, and 0.15, respectively. At this time, it can be considered that the average performance of the three is relatively good. Then, the selection is made by comparing their stability. If the stability evaluation results of different models are very close, the fluctuation ranges have a large overlap (the overlapping range is greater than 75% of any model) and the difference in fluctuation ranges is less than 10% of the optimal prediction rate. For example, the prediction rate fluctuation range of model D is 0.85-0.95, and the prediction rate fluctuation range of model F is 0.84-0.94. At this time, it can be considered that the prediction stability of the two models is not much different. In this case, a simpler model with fewer parameters is selected to ensure the actual application effect on low-performance terminals.

[0061] Step 5: Apply model prediction: Use the optimal prediction model to predict the zoning type of the target sandstone-type uranium ore sample image and visualize the prediction results. Figure 5 The prediction results and predicted zoning probability diagram of different zoning samples of sandstone-type uranium deposits using the optimal model provided in this embodiment.

[0062] Furthermore, in one embodiment of the present invention, the optimal model training at the S4 evaluation point is used to predict the zoning type and grain size of the sandstone-type uranium ore core, and the prediction results are visualized, including:

[0063] Prediction process: Use the optimal prediction model to predict the zoning type of the target sandstone-type uranium ore sample image and visualize the prediction results. First, the target sandstone-type uranium ore sample image is preprocessed with the same operations as the training data (including image format conversion, denoising, repair, etc.) to make it meet the requirements of the model input.

[0064] Then, the preprocessed image is input into the optimal prediction model, and the model outputs the predicted band type based on the relationship between the learned image features and the band and granularity labels.

[0065] Visualization: The predicted zoning labels corresponding to the reservoir depth are visualized on the image. For example, the zoning type (such as "ore-bearing section", "reduction zone", "oxidation zone") can be marked in a corner of the image, or different zoning areas can be marked with different colors to intuitively display the prediction results and provide accurate zoning information for the exploration and development of sandstone-type uranium deposits.

[0066] The typical interlayer oxidation zone identification device of sandstone-type uranium ore of the present invention is composed of an acquisition module, a preprocessing module, a training module, a verification module and a prediction module, and each module works together to realize the zoning of sandstone-type uranium ore and the intelligent identification of sandstone particle size.

[0067] The functions of each module are as follows:

[0068] Acquisition module: responsible for collecting images and recording zoning and grain size labels, manually matching to form a data set and automatically normalizing. Use digital cameras and other equipment to collect cross-sectional images of sandstone-type uranium ore cores, and based on the comparison of geological survey data and the experience of relevant professionals, determine zoning and grain size information, match image labels, and form an initial data set.

[0069] Preprocessing module: Use digital image technology and data preprocessing methods to process images in the data set. After unifying the image format to JPG or PNG, use the Faster RCNN model to segment the core area and remove irrelevant interference such as background; then cut to 256*256 resolution to meet the input requirements of the model; then use denoising and crack repair technology to improve image quality; finally, use image data enhancement operations such as flipping, rotation, and brightness adjustment to expand data diversity and improve model generalization capabilities.

[0070] Training module: Select different models in the image classification toolbox, write or modify configuration files, and call pre-trained models for deep learning training. Use pre-processed core image data to drive model deep learning training.

[0071] Verification module: Evaluate the model training results, select models with sufficiently high TOP1 prediction rates and low loss function values, perform cross-validation, and determine the optimal model after multiple rounds of training and verification cycles.

[0072] Prediction module: Use the optimal prediction model to predict the zoning of sandstone-type uranium deposit core images and visualize the prediction results.

[0073] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for identifying typical interlayer oxidation zones in sandstone-type uranium deposits, characterized in that: The following steps are involved: Step 1: Obtain a core cross-section image of a typical interlayer oxidation zone sandstone-type uranium deposit, compare the core cross-section image with the spatial position information of the sample, and determine the stratigraphic layer information to which the sample belongs, and then determine the zoning label of the sample; determine the sandstone grain size by observing the grain size and distribution characteristics of the sandstone core, and produce a sample data set containing zoning labels and grain size labels; Step 2: Preprocess the images in the sample data set using digital image technology, and divide the sample data set into a training set and a test set; Step 3: Build and train the model: Select a variety of different pre-trained neural network models for visual classification and use the pre-processed training set to start the deep learning training of the model; Step 4: Evaluate and verify the model: Evaluate the model training results and select the optimal prediction model; Step 5: Apply model prediction: Use the optimal prediction model to predict the zoning type of the target sandstone-type uranium ore sample image and visualize the prediction results.

2. The method for identifying typical interlayer oxidation zones of sandstone-type uranium deposits according to claim 1, characterized in that: The samples obtained in step 1 should cover image contents of different zones and different particle sizes, including the following five categories: mineral-bearing coarse sandstone, mineral-bearing fine sandstone, reduced zone fine sandstone, oxidized zone fine sandstone, and oxidized zone coarse sandstone.

3. The method for identifying typical interlayer oxidation zones of sandstone-type uranium deposits according to claim 2, characterized in that: In step 1, the screening and processing requirements for the core cross-section image are as follows: the captured image must ensure accurate focus, clear shooting, appropriate brightness, and the core must not be contaminated by impurities.

4. The method for identifying typical interlayer oxidation zones of sandstone-type uranium deposits according to claim 1, characterized in that: The specific implementation method of the image preprocessing in step 2 is: Step 21: Convert the image formats to JPG or PNG. Step 22: Core image segmentation: For each image in the sample dataset, use the Faster RCNN image segmentation model to extract the core part from the image; Step 23, image cutting: cutting the extracted image containing only the core part into an image with a resolution of 256*256; Step 24: Use denoising technology to remove noise interference in the image; Step 25: Repair the cracks in the image to restore the integrity of the image; Step 26, perform image enhancement: image flipping, rotation, and brightness adjustment; Step 27, data set division: divide the sample data set after image enhancement into training set and test set.

5. The method for identifying typical interlayer oxidation zones of sandstone-type uranium deposits according to claim 1, characterized in that: In step 3, the pre-trained visual classification neural network models include ResNet, ResNeXt, MobileNet, DenseNet, EfficientNet, MViT, VGG and GoogleNet.

6. The method for identifying typical interlayer oxidation zones of sandstone-type uranium deposits according to claim 1, characterized in that: The specific implementation method of step 4 is: Step 41, determine the evaluation indicators: use TOP1 prediction rate accuracy and loss function loss as the evaluation indicators of the model; Step 42, model selection: input the test set data into different visual classification neural network models, calculate the TOP1 prediction rate and loss function of each model, and select the model whose TOP1 prediction rate is stably greater than 0.8 and whose loss function is stably less than 0.3 as the number of training rounds increases for cross-validation; Step 43: Cross-validation: Use K-fold cross-validation method to verify the generalization ability of the model and select the optimal model.

7. A device for identifying typical interlayer oxidation zones of sandstone-type uranium deposits, which uses the method described in any one of claims 1 to 6 to achieve the identification of typical interlayer oxidation zones of sandstone-type uranium deposits, characterized in that: Includes the following modules: Acquisition module: responsible for collecting images and recording banding and granularity labels to form a data set; Preprocessing module: Use digital image technology and data preprocessing methods to process images in the data set; Training module: select different models, write or modify configuration files, and call pre-trained models for deep learning training; Verification module: evaluates the model training results, selects models with sufficiently high TOP1 prediction rates and low loss function values, performs cross-validation, and determines the optimal model after multiple rounds of training and verification cycles; Prediction module: Use the optimal prediction model to predict the sandstone-type uranium zoning results and visualize the prediction results.

Images